From Imaging to Prognosis: Crafting Clinical Nomograms Based on a Multi-Sequence MRI Radiomics Model for Non- Invasive Glioma Survival Prediction

doi:10.21203/rs.3.rs-4721285/v1

Download PDF

Research Article

From Imaging to Prognosis: Crafting Clinical Nomograms Based on a Multi-Sequence MRI Radiomics Model for Non- Invasive Glioma Survival Prediction

https://doi.org/10.21203/rs.3.rs-4721285/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

High incidence and malignancy call for non-invasive pre-surgery survival prediction in gliomas. Radiomics serves as a mature solution bridging this gap.

Methods

We retrospectively collected preoperative MRI from 353 patients with diffuse gliomas, comprising 108 from our institution (Center1) and 137 from The Cancer Genome Atlas dataset (TCGA) as the training cohort, with an external 108 cases from Center1 serving as an independent test cohort. Radiomic features were automatically extracted from MRI, including structural image of T1WI, T2WI, T1CE, FLAIR, and functional image of DWI (b = 1000), and ADC maps. Following a series of feature engineering and machine learning procedures, features were selected to construct the prognostic model, culminating in the radiomics survival biomarker (RadSurv). The efficacy of individual structural and functional sequences and their combinations were evaluated at all glioma, glioblastoma (GBM) and non-GBM levels using the concordance index (C-index). The optimal all-sequence combination model's RadSurv then underwent Kaplan-Meyer analysis and Cox regression analysis, and we finally developed nomograms.

Results

T1CE, ADC and FLAIR led single-sequence performance, while DWI lagged. T1CE was optimal for all glioma (C-index = 0.799) and GBM (C-index = 0.706), whereas ADC for non-GBM (C-index = 0.917). Multi-sequence combinations didn't improve predictions (C-index _Glioma = 0.787, C-index _GBM = 0.689, C-index _non−GBM = 0.893), but them win over adaptability. RadSurv effectively stratified risk for the all glioma, GBM and non-GBM populations through three predetermined cut-off values. Multivariate Cox regression confirmed RadSurv as an independent prognostic factor. The nomogram, constructed from RadSurv and age, accurately predicted survival probabilities and median survival times for glioma patients at various time points, especially for GBM.

Conclusions

The preoperative radiomics model's prognostic biomarker, RadSurv, effectively stratifies risk in glioma patients and, through nomograms, enables precise and quantifiable predictions of patient survival outcomes, warranting its utilization in clinical practice.

Glioma

Radiomics

Survival analysis

Prediction model

Nomogram.

Gliomas are the most prevalent primary brain tumors in the central nervous system (CNS). Despite aggressive treatments, including maximal safe resection and concomitant chemoradiotherapy[1], the prognosis for these patients remains dismal. According to the fifth edition of the WHO Classification of CNS tumors, published in 2021, adult diffuse gliomas are categorized into astrocytoma, oligodendroglioma and glioblastoma (GBM) in the integration of histopathological and molecular information[2][3]. Based on epidemiologic investigations[4][5][6][7], astrocytomas demonstrate 1- and 3-year survival rates of around 70% and 40%, respectively, with a median survival of 3–10 years. Oligodendrogliomas show higher 1- and 3-year survival rates of roughly 90% and 70%, and individuals often survive for more than 10 years. In stark contrast, GBMs, known for their aggressive invasiveness and rapid proliferation, have much lower survival rates—approximately 40% after one year and only 5% after three years, with a median survival of only 12–18 months. For neurosurgeons, whether in the outpatient clinic, pre-surgery, or post-surgery, the primary concern of patients and their families is always, "How long can I live with this disease?", and this information holds particular significance in the preoperative phase as it aids surgeons in formulating an appropriate treatment plan.

MRI currently holds a significant position for the diagnosis and prognostic assessment of gliomas[8]. Through established radiomics techniques, which extract high-dimensional feature matrices from images[9], researchers can assess the tumor risk level by evaluating signal intensity, shape characteristics, and texture features (e.g., image homogeneity, edges, and roughness) within the lesion area, thus empowering doctors to gain more precise prognostic insights of patients. Previous studies have primarily focused on structural MRI, namely T1-weighted imaging (T1WI), T1 contrast-enhanced weighted imaging (T1CE), T2-weighted imaging (T2WI), fluid attenuated inversion recovery (FLAIR), to investigate these issues[10][11][12]. Some have suggested that radiomics models based on [18F]FET PET imaging can help predict short-term survival (< 12 months) in newly diagnosed GBM[13] or detect glioma infiltration[14]. However, the heightened financial burdens imposed on patients limit the sample size and pose implementation challenges. Diffusion-weighted imaging (DWI), along with the resulting apparent diffusion coefficient (ADC) maps, have become standard in the clinical management of gliomas due to their ease of acquisition. Several studies [14][15][16] have noted the potential of these sequences in predicting survival among glioma patients. However, a comprehensive radiomics assessment and comparison of structural images, DWI, and ADC maps have yet to be conducted to fully understand their impact on survival predictions.

In this study, we assess the prognostic impact of radiomics models that utilize multicenter datasets and various structural MRI sequences. Apart from analyzing individual sequences (T1WI, T1CE, T2WI, FLAIR, DWI, and ADC map), we have constructed radiomics models using a combined features from the first four structural sequences and another using all six sequences. Our goal is to develop a preoperative, non-invasive model that can predict the overall survival (OS) time and rate for glioma patients. To assist clinicians with this task, we have created a supportive nomogram.

2.1 Patient enrolment

We retrospectively collected 353 diffuse glioma patients across three datasets: 216 from our institution (Center1) and 137 from the Cancer Genome Atlas (TCGA). The inclusion criteria were as follows: (1) patients over 18 years at initial diagnosis with pathologically confirmed primary diffuse glioma post-surgery or biopsy; (2) availability of molecular information that support the WHO-CNS5 classification; (3) postoperative survival exceeding 4 weeks, with survival data accessible; (4) preoperative MRI with all sequences necessary for this study; (5) exclusion of cases with image quality compromised by factors such as patient movement or significant artifacts; (6) exclusion of patients who had undergone radiotherapy or chemotherapy prior to the MRI; (7) exclusion of cases where the tumor was located in the brainstem and deemed inoperable. The detailed inclusion and exclusion data are depicted in Supplementary Fig. 1. We divided the patients into a training cohort composed of half of the Center1 and all TCGA cases, and a test cohort consisting of the remaining Center1 cases. Baseline characteristics of all datasets are presented in Table 1.

Table 1

Datasets used in this work within the training cohort and test cohort. Tumor lesions involving multiple brain regions have been counted multiple times. SD = Standard Deviation.
	Training cohort		Test cohort
	Center1 (N = 108)	TCGA (N = 137)	Center1 (N = 108)
Age (year)
Mean (SD)	53.8 (13.8)	54.4 (14.0)	51.2 (13.3)
Median [Min, Max]	55.5 [20.0, 84.0]	55.0 [21.8, 85.7]	52.0 [25.0, 82.0]
Sex
Male	65 (60.2%)	71 (51.8%)	57 (52.8%)
Female	43 (39.8%)	66 (48.2%)	51 (47.2%)
Status
Alive	52 (48.1%)	51 (37.2%)	61 (56.5%)
Dead	56 (51.9%)	86 (62.8%)	47 (43.5%)
Survival time (month)
Mean (SD)	25.4 (20.7)	21.5 (17.5)	28.6 (22.0)
Median [Min, Max]	19.1 [3.17, 97.8]	16.9 [2.13, 95.6]	23.6 [2.13, 97.8]
CNS5 class
Astrocytoma	19 (17.6%)	16 (11.7%)	21 (19.4%)
Oligodendroglioma	20 (18.5%)	38 (27.7%)	19 (17.6%)
Glioblastoma	69 (63.9%)	83 60.6%)	68 (63.0%)
IDH status
Mutation	39 (36.1%)	54 (39.4%)	40 (37.0%)
Wild-type	69 (63.9%)	83 (60.6%)	68 (63.0%)
1p19q status
Co-deletion	22 (20.4%)	50 (36.5%)	22 (20.4%)
Non Co-deletion	56 (51.9%)	16 (11.7%)	61 (56.5%)
Not available	30 (27.8%)	71 (51.8%)	25 (23.1%)
Region
Left (vs Right)	54 (50.0%)	74 (54.0%)	56 (51.9%)
Midline crossing	7(6.5%)	6 (4.4%)	9 (8.3%)
Frontal lobe	52 (48.1%)	55 (40.1%)	59 (54.6%)
Temporal lobe	39 (36.1%)	64 (46.7%)	32 (29.6%)
Parietal lobe	33 (30.6%)	38 (27.7%)	23 (21.3%)
Occipital lobe	15 (13.9%)	26 (19.0%)	8 (7.4%)
Insular lobe	42 (38.9%)	54 (39.4%)	43 (39.8%)

Patients underwent comprehensive preoperative planning and tumor resection under neuro-navigation, with biannual telephone or outpatient follow-ups post-discharge. Baseline and survival data for TCGA dataset was acquired from the GlioVis website (http://gliovis.bioinfo.cnio.es/ ). All used MRI data used were anonymized, and the study was approved by the ethics committee of XXX (anonymized for review). Given the exponential growth in radiomics research, standardized imaging study methodologies are necessary. Our research complies with the CLEAR checklist[17], detailed in Supplementary Method 1.

2.2 MRI data acquisition and preprocessing

MRI data from the TCGA datasets can be accessed through the TCIA website (http://www.cancerimagingarchive.net/ ), including scans from 1.5T or 3.0T MRI scanners with axial sections. At our institution, preoperative imaging was done using 3.0T-MRI scanners (Siemens Magnetom Verio and Skyra) equipped with eight-channel head coils. Each patient underwent standard axial T1WI, T2WI, and FLAIR sequences, followed by contrast-enhanced, sagittal, thin-slice 3D-T1WI following an intravenous bolus of gadopentetate dimeglumine (Gd-DTPA) at 0.1 mmol/kg and 4 ml/s. We performed DWI acquisition before injecting the contrast agent using a spin-echo echo planar imaging (SE-EPI) sequence applying diffusion gradients in three orthogonal directions to reduce anisotropy effects in the DWI signal. The scanning parameters for the original sequences are detailed in Supplementary Table 1. All sequence files were initially converted from DICOM to NIFTI format using dcm2niix (https://github.com/rordenlab/dcm2niix ), followed by extraction of three-dimensional images at b-values of 0 s/mm² and 1000 s/mm² for the DWI data using the "SimpleITK" library in Python. ADC values for all voxels were calculated using a mono-exponential model [18]: ADC = ln[S(b)/S(0)]/b, where S(b) represents the signal intensity at a given b-value, and S(0) is the signal intensity at b = 0 s/mm², resulting in the generation of the ADC map. The DWI with a b-value of 1000 s/mm² was selected for subsequent processing.

After obtaining six 3D NIFTI files, we performed MRI preprocessing that included: (1) Rigid registration of all sequences with the average brain template atlas of MNI152 using Advanced Normalization Tools (ANTs, https://github.com/ANTsX/ANTs ), thereby adjusting spatial positioning and resampling the sequences to an MNI space (https://nist.mni.mcgill.ca/atlases/ ) with dimensions of 193×229×229 and voxel size of 1mm×1mm×1mm. (2) Each of the remaining five sequences underwent rigid registration to the T1CE sequence to ensure alignment within the image set of the same patient, designating T1CE as the reference for further non-linear registration processes. (3) Abandoning the strategy of individually registering each sequence to the standard brain atlas, we built upon the foundation of the rigid registrations by subjecting the T1CE sequence to further affine registration and the concluding SyN registration to the MNI152 template. During this process, ANTs generated intermediary files containing transformation parameters, which were then applied to the remaining five sequences, thus enhancing computational efficiency while maintaining alignment with the T1CE sequence. (4) All sequences underwent N4 bias field correction [19] to improve image quality. (5) The "White Stripe" R package [20] was utilized for normalization of image signal intensity. (6) The brain was extracted using the corresponding MNI152 brain mask, culminating in the acquisition of fully preprocessed images. A schematic representation of these steps is illustrated in Fig. 1A.

2.4 VOI segmentation and feature extraction

Utilizing Lin et al.'s open-source CKD-TransBTS framework [21], we performed fully automated glioma segmentation. This model synergistically harnesses the advantages of Transformers and CNNs, affording the capacity for precise localization of lesion boundaries in 3D imaging. It outperformed others in the BraTS 2021 challenge, achieved state-of-the-art (SOTA) performance. The model evaluates features on T1WI, T2WI, FLAIR, and T1CE sequences to generate volumetric region of interest (VOI), identifying three areas: enhancing tumor (ET), peritumoral edema (ED), and necrotic core (NCR).

Subsequently, the segmentation outputs were subjected to the review of two neuroradiologists with over a decade of professional experience. They imported the preprocessed sequences and the VOIs into 3DSlicer (https://www.slicer.org ), meticulously verifying the congruence of the ET, ED, and NCR regions with their radiological significance. Utilizing semi-automatic algorithms such as thresholding and flood-filling methods, they refined the segmentation and eliminated erroneously delineated areas. This process ensured the segmentation's precision while maintaining high efficiency. An illustrative example of the delineation results is presented in Fig. 1A.

As a significant number of non-GBM cases in the dataset did not show post-contrast enhancement or an ET area, we redefined these subregions into two clinically relevant areas: tumor core (TC), which includes ET and NCR, and whole tumor (WT), comprising both ED and TC. After this reclassification, we extracted features using the "PyRadiomics" library[22] in Python, employing default settings. The feature set incorporated 18 first-order features, 14 shape and size features, and 75 textural features. This was further expanded by 744 wavelet, 372 LBP, and 186 LoG filtered features, plus an additional 93 features each from exponential, gradient, logarithm, square, and square root filters for each TC and WT region on individual MRI sequences. Consequently, for each case, across all six sequences, we extracted a total of 22,488 features (1874 features for each TC and WT region multiplied by 6 sequences and by 2 regions).

2.5 Preliminary Screening of Radiomics Features

Given the generation of numerous features, feature engineering was critical in this study. Initially, we preprocessed the 1874 features extracted from each VOI subregion—which included both continuous and categorical radiomics data—by discarding features where a single value constituted over 75% of the data. This step reduced the total feature count from 22,488 to 20,818. Next, we applied the intraclass correlation coefficient (ICC) to further exclude features with instability across different radiologist interpretations, retaining 8,243 features at an ICC2k threshold of greater than 0.9. Since the fourth-order radiomics features, which were derived from various filters and parameters based on the first three orders (first-order, shape and size, textural features), resulted in a plethora of approximate features, we reduced redundancy within the TC and WT feature sets across all six sequences to mitigate multicollinearity in our analysis. This was achieved by identifying pairs of highly correlated features (Pearson correlation, r > 0.95, p < 0.05) and eliminating the one with the higher variance inflation factor (VIF) from each pair, culminating in the retention of 3,543 features. Figure 1B illustrates the distributions of the resulting feature set and the ICC filtering process.

2.6 Machine-learning-based prognostic prediction Feature selection and model construction

Within the training cohort, we established a robust feature selection and modeling process for prognosis using machine learning algorithms. Input features were divided into eight groups, each yielding a model and a final radiomics survival biomarker (RadSurv). The groups included: 510 features for T1WI, 584 for T2WI, 686 for T1CE, 591 for FLAIR, 565 for DWI, 606 for ADC, 2,372 for the combined first four sequences (Combo-4), and 3,543 for all six sequences combined (Combo-6). Subjects were randomly split between the training and validation sets in ratios of 3:7, 4:6, 5:5, 6:4, or 7:3 across 500 iterations, with each iteration following a specified set of steps.

First, within the training set, univariate Cox regression was employed for selecting prognostic features (likelihood ratio test, p < 0.05). Next, we calculated risk scores (risk score = Σ feature value × feature's Cox coefficient) to stratify the training set into high- and low-risk groups, using the "survminer" package's "surv-cutpoint" function to find the optimal cut-off point. If a significant difference between the groups was observed in the Kaplan-Meyer analysis (log-rank test, p < 0.05), these features were retained for the validation set; otherwise, the process iterated further. Notably, to enhance the diversity and efficiency of the prognostic feature selection methodology, following initial univariate Cox regression, we incorporated two additional independent processes: a Cox regression model with Least Absolute Shrinkage and Selection Operator (Cox-LASSO) and a Random Survival Forest Variable Hunting (RSFVH) algorithm. We refer to this basic approach, which did not include either of the two algorithms, as Cox-regression-based Feature Selection (Cox-FS). In the validation set, each radiomics features identified from the previous training set step was further evaluated for prognostic value across four tasks: (i) value in GBM, (ii) value in non-GBM, (iii) value in all glioma cases, and (iv) value concurrently in GBM and non-GBM. The evaluation criteria were consistent with those used in the training set. We tracked how often each feature passed these validations across all sampling ratios and iterations and finally generated a count table of prognostic efficacy, one for each feature selection method across eight scenarios. These steps are outlined in Fig. 1C.

We normalized the count tables to mitigate disparities in absolute values before using unsupervised clustering to distill prognostically potent feature clusters from the count data. Approximately two-thirds or more of the features are expected to be culled through this selection mechanism. The clustering results are depicted in Supplementary Fig. 2. The 95% threshold previously set for multicollinearity meant effective clustering without overly dispersing features, although the retained features may still harbor some multicollinearity, potentially affecting subsequent modeling. To mitigate this issue, a further feature selection protocol was introduced to isolate highly efficacious, independent variables. Starting at a threshold of 0.9 and decreasing in 0.05 intervals to a minimum of 0.1, we pruned pairs of redundant features by removing the one with the higher VIF at each decremental threshold, as illustrated in Fig. 1D. Ultimately, in the training cohort, a multifactorial survival analysis model was constructed. This was accomplished by employing stepwise regression to further reduce potential multicollinearity or by bypassing stepwise regression and proceeding directly to modeling. The scoring formula is as follows:

RadSurv = h0(t) × exp (β1X1 + β2X2 + ... + βnXn)

The formula includes X as radiomic feature values, β as the regression coefficients, and h0(t) as the baseline hazard. This concludes the exposition of the selection and final modeling methodologies for eight scenarios of feature groups using three parallel approaches: Cox-FS, Cox-LASSO, and RSFVH. The processes described herein were executed by our custom-built automated script.

2.7 Evaluation

In this study, we thoroughly assessed RadSurv's survival prediction performance across four key aspects. Firstly, models were trained for various scenarios using the methods previously described, and Harrell's concordance index (C-index) was calculated with the "lifelines" library in python to quantitatively evaluate the predictive power for glioma prognosis. The results were represented in Table 2, Fig. 2 and Supplementary Table 2. We focused on comparing three feature selection methods (Cox-FS, Cox-LASSO, RSFVH), the radiomics model's effectiveness across different tasks (all gliomas, GBM, non-GBM), and the performance between and combination of imaging sequences (Combo-4 and Combo-6). Secondly, we explored RadSurv's ability to stratify patients into risk categories, both in GBM and non-GBM subgroups, exemplified with an optimal Combo-6 model. Kaplan-Meyer curves were further plotted with the "survival" and "survminer" R packages to assess and demonstrate the risk stratification capacity at various cut-off values (Supplementary Method 2 described the method of finding the best cut-off values) for glioma (Fig. 3). Thirdly, to ascertain whether the laboriously trained and explored RadSurv was an independent prognostic factor for glioma, univariate and multivariate Cox regression analyses were performed using the "rms" R package, with results presented in Table 3. Lastly, we constructed nomograms and performed calibration curve validations (Fig. 4) using the "regplot" R package. These validations encompassed the all glioma cases and delved into GBM and non-GBM subgroups.

Table 2

The best C-index scores for structural, functional, and combined sequences with different feature selection methods for varied tasks on the training cohort. The three data framed in boxes represent the best results for the ALL, Non-GBM, and GBM groups, respectively. In the first six rows of table, data in bold italic indicate the top three performance scores within the current glioma grouping. In the last two rows, data bolded and underlined signify the highest scores in their grouping. Combo-4 = structural-sequence combinations; Combo-6 = all-sequence combinations.
	ALL			non-GBM			GBM
	CoxFS	CoxLasso	RSFVH	CoxFS	CoxLasso	RSFVH	CoxFS	CoxLasso	RSFVH
T1WI	0.705	0.732	0.698	0.835	0.893	0.826	0.598	0.643	0.606
T2WI	0.675	0.740	0.715	0.793	0.777	0.744	0.630	0.680	0.660
T1CE	0.776	0.768	0.799	0.744	0.860	0.760	0.643	0.632	0.706
FLAIR	0.738	0.726	0.719	0.909	0.876	0.835	0.664	0.635	0.630
DWI	0.703	0.690	0.723	0.851	0.769	0.769	0.631	0.602	0.615
ADC	0.716	0.703	0.701	0.868	0.901	0.917	0.634	0.600	0.605
Combo-4	0.772	0.764	0.772	0.835	0.868	0.826	0.689	0.649	0.651
Combo-6	0.753	0.755	0.787	0.893	0.835	0.893	0.633	0.644	0.683

Table 3

Cox regression analysis of prognostic factors in discovery cohort and validation cohort. Note. ^a Numerical variable, ^b Categorical variable, OL = Oligodendroglioma, GBM = Glioblastoma, AS = Astrocytoma, WT = wild-type, MT = mutant, Cod = co-deletion, NA = not available.
Variable	Training cohort				Test cohort
	Univariate analysis		Multivariate analysis		Univariate analysis		Multivariate analysis
	HR (95%CI)	p.value	HR (95%CI)	p.value	HR (95%CI)	p.value	HR (95%CI)	p.value
RadSurv^a	1.30 (1.20–1.30)	< 0.001	1.18 (1.12–1.25)	< 0.001	1.05 (1.00-1.10)	0.0017	1.08 (1.00-1.17)	0.043
Age^a	1.10 (1.10–1.10)	< 0.001	1.03 (1.01–1.05)	0.007	1.10 (1.00-1.10)	< 0.001	1.05 (1.02–1.08)	0.002
Gender^b	0.94 (0.68–1.30)	0.72			0.84 (0.47–1.50)	0.55
CNS5 class^b
OL vs GBM	0.14 (0.07–0.27)	< 0.001	0.53 (0.16–1.78)	0.301	0.06 (0.01–0.44)	0.0056	0.03 (0.00-0.60)	0.023
AS vs GBM	0.20 (0.10–0.44)	< 0.001	NA	NA	0.28 (0.11–0.72)	0.0086	NA	NA
IDH status^b
WT vs MT	9.08 (5.80–17.0)	< 0.001	3.72 (1.53–9.06)	0.004	11.0 (4.50–28.0)	< 0.001	4.72 (1.36–16.4)	0.015
1p/19q status^b
Cod vs non-cod	0.28 (0.18–0.45)	< 0.001	1.59 (0.78–3.26)	0.205	0.25 (0.09–0.69)	0.0078	4.39 (1.15–16.8)	0.031
NA vs non-cod	3.50 (2.50-5.00)	< 0.001	1.18 (0.76–1.84)	0.466	3.30 (1.70–6.10)	< 0.001	1.41 (0.72–2.80)	0.318
Location^b
Left vs right	0.87 (0.62–1.20)	0.4			0.91 (0.51–1.60)	0.76
Cross midline	0.97 (0.45–2.10)	0.94			0.81 (0.29–2.30)	0.69
Frontal lobe	0.71 (0.50-1.00)	0.051			0.62 (0.35–1.10)	0.098
Temporal lobe	1.70 (1.20–2.30)	0.0019	1.61 (1.13–2.29)	0.008	2.30 (1.30–4.10)	0.0055	1.36 (0.62–2.96)	0.444
Parietal lobe	1.30 (0.90–1.80)	0.17			2.60 (1.40-5.00)	0.0027	1.49 (0.63–3.56)	0.366
Occipital lobe	1.60 (1.10–2.40)	0.019	1.13 (0.74–1.73)	0.567	5.60 (2.40–13.0)	< 0.001	1.10 (0.30–3.98)	0.887
Insular	1.30 (0.92–1.80)	0.14			2.10 (1.20–3.70)	0.012	1.36 (0.61–3.03)	0.453

2.9 Statistical Analysis and data availability

The MRI processing and statistical analyses were conducted using Python (v3.8) and R (v4.2.3). Manual rectifications of VOIs were performed with 3DSlicer (v5.2). The visualization of figures and tables was accomplished through R, Microsoft Office Excel, and Adobe Illustrator. The C-index ranges from 0 to 1, where values near 1 indicate better prediction; 0.5 means no predictive ability, and below 0.5 suggests inaccurate predictions. Log-rank test was used to assess significance between two groups in Kaplan-Meyer survival analyses, considering p-values < 0.05 statistically significant. Closer alignment of nomogram calibration points with the diagonal shows greater predictive accuracy. Our code is open-source and available at https://github.com/shadiendo/Radiomic_PrognosicPredict. Additionally, we have provided the survival data from our study and a web-based dynamic nomogram solution using the 'DynNom' R package for easy visualization and utilization. Data from our institution will be provided upon reasonable request to the corresponding author.

3.1 Comparative Evaluation of CoxFS, CoxLASSO, and RSFVH Feature Selection Methods

The examination of the bolded or framed data results in Table 2 illustrates the edge of the RSFVH algorithm in selecting prognostic features (the best results are all trained by RSFVH), compared to those of CoxFS and CoxLasso. In previous descriptions, we incrementally mitigated multicollinearity to identify the optimal models. Dissecting the data of Supplementary Table 2 yields the line graphs in Fig. 2, which represents the VIF and C-index values of the corresponding model across different tasks. Figure 2 graphically illustrates that the dashed line, representing CoxFS, exhibits more pronounced fluctuations across sequences and between GBM and non-GBM, particularly evident in the "T1CE", "DWI”, and "combo-6" subplots. This indicates a potential for the highest performance ceiling but also a risk of poorer model outcomes. Observations indicate that for all three algorithms, C-index are unstable and generally poor when the VIF is low (shown in the far left of the subplots). This suggests that reducing collinearity too aggressively can remove key features, resulting in worse training outcomes. The best performances typically occur around a VIF(log) of 0.5 (equivalent to a VIF value of about 3).

3.2 Radiomics Model Demonstrates Superior Prognostic Efficacy for GBM Compared to non-GBM

As seen in Table 2, regardless of the algorithm or imaging sequence employed, the C-index for non-GBM are consistently higher by roughly 0.2 compared to those for GBM. For the best models, a RSFVH-based model on ADC reaches a C-index of 0.917 in non-GBM, while the top GBM model, using RSFVH on T1CE, scores a C-index of 0.706—significantly lower by 0.211. Further observation of Fig. 2 allows for the easy discernment that the blue line representing non-GBM is invariably positioned above that of GBM under all conditions. This may suggest that GBM's radiographic features are more complex and profound, which is less amenable to extraction via current radiomics techniques.

3.3 Evaluation of Models for Six Individual MRI Sequences

To compare the prognostic efficacy of different sequences, we selected the average of the best performance for each individual sequence in all glioma, GBM, and non-GBM showed in Table 2. The results were C-indexes of 0.788 for T1CE, 0.770 for FLAIR, 0.756 for ADC, 0.756 for T1WI, 0.738 for T2WI, and 0.735 for DWI. It is evident that T1CE and FLAIR exhibit superior performance across various sequences. Moreover, we observe that T1CE delivers the most robust results for all gliomas, including GBMs, while ADC and FLAIR proves particularly effective for non-GBM gliomas. DWI, however, ranks as the least effective (The trends are clearly seen in Fig. 2, especially when compared to Combo-6). Despite T1CE's high C-index, the original scans in our datasets are typically thinner, intrinsically containing more image details compared to other sequences that need interpolation in the resampling step to meet MNI space standards. This is precisely why T1CE is designated as the benchmark sequence for multi-sequence MRI registration with standard brain. Therefore, it would be inequitable to conclude that T1CE is unequivocally superior to other sequences without further comparative trials under identical original size and voxel parameters.

3.4 Evaluation of Models for Two Multi-MRI Sequences

Models build on individual sequences exhibit significant limitations. For example, although T1CE predicts well for GBM patients’ survival, it performs poorly for LGG compared to other sequences, such as ADC. Meanwhile, ADC predicts LGG exceptionally well but performs very poorly in GBM tasks. Therefore, a combination of different sequence features is necessary. We're keen to see if combining features from multiple sequences really pays off for better survival prediction efficacy after considerable effort.

By calculating the average of the best C-index values in three tasks in Table 2, we found that both Combo-4 (C-index = 0.776) and Combo-6 (C-index = 0.788) performed very well. However, when it comes to the three specific tasks, the performance of the combined sequence models is inferior to that of the individual sequences. For example, Combo-6 is weaker than T1CE in all gliomas (C-index 0.787 vs 0.799), weaker than ADC in non-GBM (C-index 0.893 vs 0.917), and weaker than T1CE in GBM (C-index 0.683 vs 0.706). This pattern is consistent not only in the best-performing models showed in Table 2 but also across the top five and top three models (Supplementary Table 2), bolstering these outcomes aren't random.

Overall, models based on multi-sequence features do not significantly outstrip their single sequence counterparts in effectiveness, but due to its lack of obvious shortcomings, it is suitable for clinical prediction tasks.

3.5 Leveraging the Optimal Combo-4 Model for Glioma Survival Prediction

Furthermore, taking the best model from Combo-6 (C-index = 0.787, see Table 2) as an example (Kaplan Meyer analysis, Cox regression, and Nomogram all use this example, and will not be reiterated thereafter), we showed its feature composition in Fig. 3A. This model has ten low-correlation features (average VIF of 2.33), indicating a diverse set and representative composition. Of these, ADC contributes three features, T1CE five, and T2WI two, highlighting T1CE's significant role, with no features from DWI. From an empirical perspective, the first two sequences may be due to their features being replaced by T1CE and T2WI, respectively, while DWI is simply ineffective, resonating with the visual evidence in Table 2. Only two of the ten features are TC, with the remaining eight being WT, indicating a higher utility of WT features. Additionally, the features include one shape-related (RF8), one first-order intensity feature (RF2), and four wavelet-transformed features (RF1, RF3, RF9, RF10), with the rest derived from different filters.

Among these features, RF8 contributes the most (reaching 39.9%), indicating that the more irregular the shape of the tumor, the worse the prognosis. RF7 and RF9 also play substantial roles, suggesting that the more disordered the local texture features, the poorer the prognosis. Notably, RF2 indicates that higher ADC signal values are also associated with worse outcomes. However, although contributions are quantified in Fig. 3A, the actual RadSurv result for each individual patient is also influenced by the absolute value of certain features, particularly for those with a skewed distribution, such as RF1 and RF6. When these feature values are exceptionally high in some cases, their impact on the overall RadSurv score is more significant.

3.6 Risk Stratification Proficiency of RadSurv

As depicted in Fig. 3B, we conducted Kaplan-Meyer analyses on the model, upon establishing stable cut-off values within the training cohort. We tested RadSurv's ability to distinguish between high and low-risk glioma patients, then specifically in GBM and non-GBM groups. The model effectively stratified patients based on consistent cut-off values of different subgroups, showing statistically significant in both training and test cohorts. Among Center1 and TCGA datasets, the first column shows a distinct separation between high and low-risk groups. In the second column, the Center1 non-GBM low-risk subgroup showed significantly longer survival, compared to TCGA. For GBM, shown in the third column, both risk groups in all datasets had much lower median survival times than non-GBM. Still, RadSurv highlighted significant differences between high and low-risk groups with a cut-off of 2.1. Low-risk GBM patients had median survivals of 24.0 months (Center1 training cohort), 17.9 months (TCGA cohort), and 36.6 months (Center1 test cohort). High-risk GBM patients had median survivals of 13.2 months (Center1 training cohort), 12.3 months (TCGA cohort), and 12.1 months (Center1 test cohort), indicating a slightly better outcome for Center1 but confirming GBM's overall poor prognosis, particularly in high-risk groups.

3.7 Model Validation on Cox Regression Analysis

To evaluate the impact of RadSurv and other factors on glioma prognosis and ascertain whether RadSurv is an independent prognostic marker, we built a univariate Cox regression model and then a multivariate model for significant factors (p < 0.05), as delineated in Table 3. For RadSurv, a logarithmic transformation was employed to normalize the data distribution. It is apparent that RadSurv (HR _training = 1.18, p < 0.001; HR _test = 1.08, p = 0.043), age (HR _training =1.03, p = 0.007; HR _test =1.05, p = 0.002), and IDH status (HR _training =3.72, p = 0.004; HR _test =7.72, p = 0.015) emerged as the most stable independent prognostic factors in both training and test cohorts. Among them, IDH status is a pivotal molecular marker within glioma research, recently established as a core determinant for differentiating glioma stem cells and is highly correlated with patient outcomes. Although the CNS5 class was promising in univariate analysis but didn't stand as an independent factor, likely impacted by IDH as all GBMs were IDH wild-type in our study. The 1p19q co-deletion didn't qualify as an independent marker, despite notable hazard ratios in univariate analysis (HR _training =0.28, HR _test =0.25), because 1p19q mutations are the molecular marker in oligodendrogliomas with better outcomes. Concerning tumor location, frontal lobe tumors tended to be protective, while temporal and occipital lobe tumors correlated with worse outcomes. Despite their significance in univariate analysis, temporal and occipital locations were not independent prognostic indicators.

3.8 Nomogram Development and Validation

To enhance the applicability of our model, nomograms were constructed based on the training cohort, with RadSurv and age to be reliable independent prognostic factors, obtained noninvasively and preoperatively. These graphical representations are exemplified by four prototypical cases in Fig. 4A, with the corresponding nomogram illustrated in Fig. 4B. Simulating the clinical application scenario, a physician, upon acquiring a patient's MRI and computing the RadSurv value, could employ the Nomogram (ALL) if the glioma subtype—GBM or non-GBM—remains unconfirmed. Once the tumor classification is determined, the Nomogram (non-GBM) or Nomogram (GBM) can be used for a more precise prognostic prediction. For instance, in the case of TCGA-19-2631, the patient, aged 74.5 years and with a RadSurv value of 13.524, approaches the GBM group's cut-off value of 2.1, and falls in the center of the Total points in both Nomogram (ALL) and Nomogram (GBM). With a survival time of 0.59 years, this case is significantly lower than GBM’s median survival, making it highly representative. By inputting the log-transformed RadSurv value of 1.131 into the Nomogram (Glioma), a score of 164 was obtained, indicating 1-year survival probabilities of 13% and 3-year and 5-year lower than 0.1%, respectively. The survival time predicted by the fitted curve was approximately 5 months (0.42 years). Given the patient's typical age of onset and the presence of ring enhancement on imaging, which suggest GBM, the Nomogram (GBM) offered better predictions: a total score of 168, with survival probabilities of 23%, 3.12%, and < 0.1%, and a similar estimated survival time of around 7 months (0.58 years). A notable shift was seen in the 3-year survival rate, dropping from 3.12% to < 0.1%, possibly due to the near-equal representation of non-GBM in the Nomogram (ALL), which could inflate the anticipated survival rate. The accuracy of these predictions is evaluated through the Nomogram calibration curves depicted in Fig. 4B, which indicate that the Nomogram (GBM) provides superior predictive performance. In contrast, both the Nomogram (Glioma) and Nomogram (non-GBM) exhibit varying degrees of over- and underestimation, particularly in the 3-year survival predictions for non-GBM patients, where the Nomogram (non-GBM) demonstrates a significant underestimation. This suggests that actual survival times in non-GBM are more variable, as seen in the scatter plots of real-life cases in the central part of each nomogram, and clinicians should consider this variability when applying these nomograms, potentially offering non-GBM patients a more optimistic outlook.

In this study, we extracted radiomics features from preoperative multi-sequence MRI to create RadSurv, a survival prognostic biomarker for gliomas. To this end, we employed C-index as a metric to meticulously assess the prognostic efficacy of different sequences and evaluate whether the combination of sequences improved our model. Furthermore, we validated the prognostic effectiveness of RadSurv using a six-sequence feature combination model through a multifaceted approach. This included Kaplan-Meyer analysis, Cox regression, and by creating nomograms that integrate clinical data to directly predict survival probabilities. The performance of RadSurv, as observed on the calibration curves, was favorable, particularly for GBM.

We trained our model using a rich dataset from a diverse array of sources, including multiple centers, scanners, and imaging protocols. The complexity of the samples contributed to the model's ability to learn more comprehensive and generalizable features pertinent to prognosis during the training phase. This strategy prevented overfitting to a single dataset. However, given the high variety in our training samples, and considering that, unlike CT, the absolute values of MRI signals lack biological meaning, image signal normalization becomes crucial. Different intensity normalization methods yield disparate results across various sequences [23]. As Salome et al. found that white stripe normalization significantly enhanced the prognostic efficacy of the T1CE sequence for GBM's survival prediction, but not for others where Z-score, ComBat and Fuzzy C-Means algorithms were identified better for T1WI, T2WI, and FLAIR, respectively[23]. Therefore, it may be necessary to apply different signal normalization methods for different sequences in the context of MRI radiomics. Currently, the Image Biomarker Standardization Initiative (IBSI) lacks specific guidelines on the appropriate selection of intensity normalization techniques [24]. To conclude, to ensure the trained model performs well and generalizes on test data, we standardized the imaging data processing pipeline in this study to balance differences among samples. Furthermore, we made numerous efforts including building stable prognostic feature selection and modeling processes to avoid methodological flaws and ultimately strengthen their validity.

Within individual MRI sequences, Studies have shown that T1CE, T2WI, and FLAIR are particularly good at predicting glioma patient survival [23][25]. MRI allows us to correlate macroscopic clinical manifestations with phenotypic differences at the cellular and even genetic levels [26][27]. The enhanced region in T1CE is associated with the disruption of the blood-brain barrier, while T2WI/FLAIR capture the proliferative tumor margins and vasogenic edema [28]. These imaging findings are consistent with substantial evidence of tumor cell infiltration in edematous regions and continued invasion along white matter tracts into the surrounding brain tissue [29][30]. Radiogenomic research by Beig et al. has demonstrated that the areas of enhancement and edema can predict the hypoxic state of GBM [31], with hypoxia being a byproduct of tumor vascular abnormalities and a critical pathway in tumor physiology that leads to therapeutic resistance and ultimately poor prognosis [32]. Our prior studies have shown that preoperative T2 radiomics prognostic biomarkers are significantly associated with glioma macrophage infiltration, particularly with M2 macrophages [33], suggesting the potential of imaging to guide immunotherapy. Beyond structural imaging, DWI and ADC are integral to modern imaging sequences. Their utility extends beyond the diagnosis of acute cerebral infarction [34] and plays a unique role in brain tumor research. Several studies have demonstrated a correlation between ADC values and tumor cell density, with higher cellularity contributing to more restricted diffusion, manifesting as lower ADC values on imaging [35][36]. Factors influencing ADC include intratumoral microscopic changes, vascular edema, cystic alterations, and necrosis [37]. Recent studies have also employed DWI and ADC in automated segmentation tasks for gliomas [38] and the combination of ^[18F]FET PET with ADC maps has been found to enhance the imaging of glioma cell infiltration zones [14]. Similarly, the mismatch between T2WI and Flair sequences helps in identifying IDH mutations [39], indicating that the integration of sequences could enhance the capability for related tasks [11][12]. In our study, the efficacy of ADC has been validated, whereas DWI's reliability is less consistent, possibly due to spatial distortion on its lower image resolution or because DWI features encapsulate less prognostic information, a hypothesis that warrants further investigation.

Our study's primary limitation is the lack of data on key survival factors like chemotherapy and tumor resection extent[8], potentially causing bias. Treatments such as radiation therapy, tumor-treating fields, or other therapeutic interventions, as well as patient ethnicity, could also contribute to survival disparities [40][41]. In addition to the aforementioned factors, integrating clinicopathological features with radiomics, including IDH status, MGMT promoter methylation, TERT promoter mutations, and even genomic alterations [42][43], has been shown to enhance prognostic accuracy. This aligns with our previous work [33][44] where non-invasive imaging predicted these features, enabling a better preoperative assessment of overall patient survival. Future research could incorporate a broader range of modalities, including assessments of cognitive impairment [45], extraction of features using NLP from medical records [46], and even intraoperative ultrasound radiomics [47], to develop a more comprehensive and integrated clinical picture. Nomograms are a focal point and a critical component of our study's presentation. The very inception of the nomogram was to provide a visual representation of complex computational formulas, facilitating a rapid, intuitive, and precise elucidation of the relationships between various variables [48]. However, incorporating an increasing number of parameters tends to decrease the model's practicality and its capacity for clinical adoption. This challenge may necessitate the development of an intelligent, automated platform capable of accommodating these multi-sequence features. Lastly, the limited amount of data in our study compelled us to used half of our center's dataset for independent testing, which remains an internal test. Future work is anticipated to conduct external tests across multiple centers.

In short, we've created a robust survival prognostic model for glioma based on our own and public datasets. Notably, in our study, all patients were classified into two categories: GBM and non-GBM. Although they have significant survival differences, it's hard to tell them apart with imaging before biopsy. Considering that the prognosis of GBM is markedly poorer compared to non-GBM, we have built a model that caters to the all glioma population and subsequently developed refined models focused on subtypes. This model is more aligned with clinical application scenarios and enables patients and their families to have a clearer understanding of the potential prognoses associated with different glioma subtypes before surgery.

TCGA

The Cancer Genome Atlas dataset

RadSurv

Radiomics Survival Biomarker

GBM

Glioblastoma

C-index

Concordance Index

T1WI

T1-Weighted Imaging

T2WI

T2-Weighted Imaging

T1CE

T1 Contrast-Enhanced Weighted Imaging

FLAIR

Fluid Attenuated Inversion Recovery

DWI

Diffusion Weighted Imaging

ADC

Apparent Diffusion Coefficient

ANTs

Advanced Normalization Tools

VOI

Region Of Interest

Enhancing Tumor

Peritumoral Edema

NCR

Necrotic Core

Tumor Core

Whole Tumor

ICC

Intraclass Correlation Coefficient

VIF

Variance Inflation Factor

Combo-4

4 Sequences Combined

Combo-6

6 Sequences Combined

RSFVH

Random Survival Forest Variable Hunting

Cox-LASSO

Least Absolute Shrinkage and Selection Operator

Cox-FS

Cox-regression-based Feature Selection

Ethics approval and consent to participate

All used MRI data used were anonymized, and the study was approved by the ethics committee of Nanjing Medical University (No. 2021-SR-076).

Consent for publication

Not applicable

Availability of data and materials

The TCGA datasets used and analysed during the current study are available in the TCIA repository (http://www.cancerimagingarchive.net/ ) and GlioVis website (http://gliovis.bioinfo.cnio.es/ ). The Center1 datasets used and analysed during the current study are available from the corresponding author on reasonable request.

Competing interests

The authors declare that they have no competing interests

Funding

This work was supported by grants from the National Natural Science Foundation of China (82372898), Jiangsu Province Capability Improvement Project through Science, Technology and Education (ZDXK202225), and the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD) and Jiangsu Province Hospital Clinical Capacity Enhancement Project (JSPH-MC-2021-27).

Authors' contributions

Xiao Fan: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Investigation, Data curation; Hongjian Zhang: Writing – review & editing, Validation, Software, Investigation; Bin Huang: Writing – review & editing, Investigation, Conceptualization; Jincheng Tao: Investigation, Data curation; Jintan Li: Investigation; Min Zhang: Investigation; Hang Zhang: Investigation; Xixi Hu: Investigation; Xiefeng Wang: Resources, Data curation; Yongping You: Resources, Data curation; Junxia Zhang: Resources; Hui Luo: Supervision, Resources, Data curation; Yun Yu: Supervision, Resources; Yingyi Wang: Writing – review & editing, Supervision, Resources, Data curation.

Acknowledgements

Not applicable

Ostrom QT, Cioffi G, Waite K, Kruchko C, Barnholtz-Sloan JS. (2021). CBTRUS Statistical Report: Primary Brain and Other Central Nervous System Tumors Diagnosed in the United States in 2014–2018. Neuro-oncology, 23(12 Suppl 2), iii1–iii105. https://doi.org/10.1093/neuonc/noab200.
Louis DN, Perry A, Wesseling P, Brat DJ, Cree IA, Figarella-Branger D, Hawkins C, Ng HK, Pfister SM, Reifenberger G, Soffietti R, von Deimling A, Ellison DW. The 2021 WHO Classification of Tumors of the Central Nervous System: a summary. Neurooncology. 2021;23(8):1231–51. https://doi.org/10.1093/neuonc/noab106.
Horbinski C, Berger T, Packer RJ, Wen PY. Clinical implications of the 2021 edition of the WHO classification of central nervous system tumours. Nat reviews Neurol. 2022;18(9):515–29. https://doi.org/10.1038/s41582-022-00679-w.
Gladson CL, Prayson RA, Liu WM. The pathobiology of glioma tumors. Annu Rev Pathol. 2010;5:33–50. https://doi.org/10.1146/annurev-pathol-121808-102109.
Stupp R, Brada M, van den Bent MJ, Tonn JC, Pentheroudakis G, ESMO Guidelines Working Group. High-grade glioma: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Annals oncology: official J Eur Soc Med Oncol. 2014;25(Suppl 3):iii93–101. https://doi.org/10.1093/annonc/mdu050.
Ostrom QT, Cote DJ, Ascha M, Kruchko C, Barnholtz-Sloan JS. Adult Glioma Incidence and Survival by Race or Ethnicity in the United States From 2000 to 2014. JAMA Oncol. 2018;4(9):1254–62. https://doi.org/10.1001/jamaoncol.2018.1789.
Stupp, R., Mason, W. P., van den Bent, M. J., Weller, M., Fisher, B., Taphoorn, M.J., Belanger, K., Brandes, A. A., Marosi, C., Bogdahn, U., Curschmann, J., Janzer,R. C., Ludwin, S. K., Gorlia, T., Allgeier, A., Lacombe, D., Cairncross, J. G., Eisenhauer,E., Mirimanoff, R. O., European Organisation for Research and Treatment of Cancer Brain Tumor and Radiotherapy Groups, … National Cancer Institute of Canada Clinical Trials Group (2005). Radiotherapy plus concomitant and adjuvant temozolomide for glioblastoma.The New England journal of medicine, 352(10), 987–996. https://doi.org/10.1056/NEJMoa043330.
Zhang K, Liu X, Li G, Chang X, Li S, Chen J, Zhao Z, Wang J, Jiang T, Chai R. Clinical management and survival outcomes of patients with different molecular subtypes of diffuse gliomas in China (2011–2017): a multicenter retrospective study from CGGA. Cancer biology Med. 2022;19(10):1460–76. https://doi.org/10.20892/j.issn.2095-3941.2022.0469.
Gillies RJ, Kinahan PE, Hricak H. Radiomics: Images Are More than Pictures. They Are Data Radiol. 2016;278(2):563–77. https://doi.org/10.1148/radiol.2015151169.
Kickingereder P, Burth S, Wick A, Götz M, Eidel O, Schlemmer HP, Maier-Hein KH, Wick W, Bendszus M, Radbruch A, Bonekamp D. Radiomic Profiling of Glioblastoma: Identifying an Imaging Predictor of Patient Survival with Improved Performance over Established Clinical and Radiologic Risk Models. Radiology. 2016;280(3):880–9. https://doi.org/10.1148/radiol.2016160845.
Wang J, Zheng X, Zhang J, Xue H, Wang L, Jing R, Chen S, Che F, Heng X, Li G, Xue F. An MRI-based radiomics signature as a pretreatment noninvasive predictor of overall survival and chemotherapeutic benefits in lower-grade gliomas. Eur Radiol. 2021;31(4):1785–94. https://doi.org/10.1007/s00330-020-07581-3.
Deng DB, Liao YT, Zhou JF, Cheng LN, He P, Wu SN, Wang WS, Zhou Q. Non-Invasive Prediction of Survival Time of Midline Glioma Patients Using Machine Learning on Multiparametric MRI Radiomics Features. Front Neurol. 2022;13:866274. https://doi.org/10.3389/fneur.2022.866274.
Li Z, Holzgreve A, Unterrainer LM, Ruf VC, Quach S, Bartos LM, Suchorska B, Niyazi M, Wenter V, Herms J, Bartenstein P, Tonn JC, Unterrainer M, Albert NL, Kaiser L. Combination of pre-treatment dynamic [18F]FET PET radiomics and conventional clinical parameters for the survival stratification in patients with IDH-wildtype glioblastoma. Eur J Nucl Med Mol Imaging. 2023;50(2):535–45. https://doi.org/10.1007/s00259-022-05988-2.
Verburg N, Koopman T, Yaqub MM, Hoekstra OS, Lammertsma AA, Barkhof F, Pouwels PJW, Reijneveld JC, Heimans JJ, Rozemuller AJM, Bruynzeel AME, Lagerwaard F, Vandertop WP, Boellaard R, Wesseling P, de Witt Hamer PC. Improved detection of diffuse glioma infiltration with imaging combinations: a diagnostic accuracy study. Neurooncology. 2020;22(3):412–22. https://doi.org/10.1093/neuonc/noz180.
Vajapeyam S, Brown D, Billups C, Patay Z, Vezina G, Shiroishi MS, Law M, Baxter P, Onar-Thomas A, Fangusaro JR, Dunkel IJ, Poussaint TY. Advanced ADC Histogram, Perfusion, and Permeability Metrics Show an Association with Survival and Pseudoprogression in Newly Diagnosed Diffuse Intrinsic Pontine Glioma: A Report from the Pediatric Brain Tumor Consortium. AJNR Am J Neuroradiol. 2020;41(4):718–24. https://doi.org/10.3174/ajnr.A6499.
Yoo RE, Choi SH, Kim TM, Lee SH, Park CK, Park SH, Kim IH, Yun TJ, Kim JH, Sohn CH. Independent Poor Prognostic Factors for True Progression after Radiation Therapy and Concomitant Temozolomide in Patients with Glioblastoma: Subependymal Enhancement and Low ADC Value. AJNR. Am J Neuroradiol. 2015;36(10):1846–52. https://doi.org/10.3174/ajnr.A4401.
Kocak B, Baessler B, Bakas S, Cuocolo R, Fedorov A, Maier-Hein L, Mercaldo N, Müller H, Orlhac F, Pinto Dos Santos D, Stanzione A, Ugga L, Zwanenburg A. CheckList for EvaluAtion of Radiomics research (CLEAR): a step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMII. Insights into imaging. 2023;14(1):75. https://doi.org/10.1186/s13244-023-01415-8.
Le Bihan D, Breton E, Lallemand D, Aubin ML, Vignaud J, Laval-Jeantet M. Separation of diffusion and perfusion in intravoxel incoherent motion MR imaging. Radiology. 1988;168(2):497–505. https://doi.org/10.1148/radiology.168.2.3393671.
Tustison NJ, Avants BB, Cook PA, Zheng Y, Egan A, Yushkevich PA, Gee JC. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010;29(6):1310–20. https://doi.org/10.1109/TMI.2010.2046908.
Shinohara RT, Sweeney EM, Goldsmith J, et al. Statistical normalization techniques for magnetic resonance imaging. Neuroimage Clin. 2014;6:9–19. https://doi.org/10.1016/j.nicl.2014.08.008.
Lin J, Lin J, Lu C, Chen H, Lin H, Zhao B, Shi Z, Qiu B, Pan X, Xu Z, Huang B, Liang C, Han G, Liu Z, Han C. CKD-TransBTS: Clinical Knowledge-Driven Hybrid Transformer With Modality-Correlated Cross-Attention for Brain Tumor Segmentation. IEEE Trans Med Imaging. 2023;42(8):2451–61. https://doi.org/10.1109/TMI.2023.3250474.
van Jjm G, Fedorov A, Parmar C, et al. Computational radiomics system to decode the radiographic phenotype. Cancer Res. 2017;77(21):e104–7. https://doi.org/10.1158/0008-5472.CAN-17-0339.
Salome P, Sforazzini F, Grugnara G, Kudak A, Dostal M, Herold-Mende C, Heiland S, Debus J, Abdollahi A, Knoll M. MR Intensity Normalization Methods Impact Sequence Specific Radiomics Prognostic Model Performance in Primary and Recurrent High-Grade Glioma. Cancers. 2023;15(3):965. https://doi.org/10.3390/cancers15030965.
Zwanenburg, A., Vallières, M., Abdalah, M. A., Aerts, H. J. W. L., Andrearczyk, V.,Apte, A., Ashrafinia, S., Bakas, S., Beukinga, R. J., Boellaard, R., Bogowicz, M.,Boldrini, L., Buvat, I., Cook, G. J. R., Davatzikos, C., Depeursinge, A., Desseroit,M. C., Dinapoli, N., Dinh, C. V., Echegaray, S., … Löck, S. (2020). The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology, 295(2), 328–338. https://doi.org/10.1148/radiol.2020191145.
Tan Y, Mu W, Wang XC, Yang GQ, Gillies RJ, Zhang H. Improving survival prediction of high-grade glioma via machine learning techniques based on MRI radiomic, genetic and clinical risk factors. Eur J Radiol. 2019;120:108609. https://doi.org/10.1016/j.ejrad.2019.07.010.
Xu PF, Li C, Chen YS, Li DP, Xi SY, Chen FR, Li X, Chen ZP. Radiomics-based survival risk stratification of glioblastoma is associated with different genome alteration. Comput Biol Med. 2023;159:106878. https://doi.org/10.1016/j.compbiomed.2023.106878.
Sun Q, Chen Y, Liang C, Zhao Y, Lv X, Zou Y, Yan K, Zheng H, Liang D, Li ZC. Biologic Pathways Underlying Prognostic Radiomics Phenotypes from Paired MRI and RNA Sequencing in Glioblastoma. Radiology. 2021;301(3):654–63. https://doi.org/10.1148/radiol.2021203281.
Kao HW, Chiang SW, Chung HW, Tsai FY, Chen CY. (2013). Advanced MR imaging of gliomas: an update. BioMed research international, 2013, 970586. https://doi.org/10.1155/2013/970586.
Kelly PJ, Daumas-Duport C, Kispert DB, Kall BA, Scheithauer BW, Illig JJ. Imaging-based stereotaxic serial biopsies in untreated intracranial glial neoplasms. J Neurosurg. 1987;66(6):865–74. https://doi.org/10.3171/jns.1987.66.6.0865.
Wang J, Yi L, Kang QM, Zhou J, Chen TQ, Hugnot JP, Yu SC. Glioma invasion along white matter tracts: A dilemma for neurosurgeons. Cancer Lett. 2022;526:103–11. https://doi.org/10.1016/j.canlet.2021.11.020.
Beig N, Patel J, Prasanna P, Hill V, Gupta A, Correa R, Bera K, Singh S, Partovi S, Varadan V, Ahluwalia M, Madabhushi A, Tiwari P. Radiogenomic analysis of hypoxia pathway is predictive of overall survival in Glioblastoma. Sci Rep. 2018;8(1):7. https://doi.org/10.1038/s41598-017-18310-0.
Wu, L., Wu, W., Zhang, J., Zhao, Z., Li, L., Zhu, M., Wu, M., Wu, F., Zhou, F., Du,Y., Chai, R. C., Zhang, W., Qiu, X., Liu, Q., Wang, Z., Li, J., Li, K., Chen, A.,Jiang, Y., Xiao, X., … Wang, Q. (2022). Natural Coevolution of Tumor and Immunoenvironment in Glioblastoma. Cancer discovery, 12(12), 2820–2837. https://doi.org/10.1158/2159-8290.CD-22-0196.
Fan X, Li J, Huang B, Lu H, Lu C, Pan M, Wang X, Zhang H, You Y, Wang X, Wang Q, Zhang J. Noninvasive radiomics model reveals macrophage infiltration in glioma. Cancer Lett. 2023;573:216380. https://doi.org/10.1016/j.canlet.2023.216380.
Schlaug G, Siewert B, Benfield A, Edelman RR, Warach S. Time course of the apparent diffusion coefficient (ADC) abnormality in human stroke. Neurology. 1997;49(1):113–9. https://doi.org/10.1212/wnl.49.1.113.
Surov A, Meyer HJ, Wienke A. Correlation between apparent diffusion coefficient (ADC) and cellularity is different in several tumors: a meta-analysis. Oncotarget. 2017;8(35):59492–9. https://doi.org/10.18632/oncotarget.17752.
Guo AC, Cummings TJ, Dash RC, Provenzale JM. Lymphomas and high-grade astrocytomas: comparison of water diffusibility and histologic characteristics. Radiology. 2002;224(1):177–83. https://doi.org/10.1148/radiol.2241010637.
Huang WY, Wen JB, Wu G, Yin B, Li JJ, Geng DY. Diffusion-Weighted Imaging for Predicting and Monitoring Primary Central Nervous System Lymphoma Treatment Response. AJNR Am J Neuroradiol. 2016;37(11):2010–8. https://doi.org/10.3174/ajnr.A4867.
Khorasani A, Kafieh R, Saboori M, Tavakoli MB. Glioma segmentation with DWI weighted images, conventional anatomical images, and post-contrast enhancement magnetic resonance imaging images by U-Net. Phys Eng Sci Med. 2022;45(3):925–34. https://doi.org/10.1007/s13246-022-01164-w.
Kurokawa R, Kurokawa M, Baba A, Ota Y, Kim J, Capizzano A, Srinivasan A, Moritani T. Dynamic susceptibility contrast-MRI parameters, ADC values, and the T2-FLAIR mismatch sign are useful to differentiate between H3-mutant and H3-wild-type high-grade midline glioma. Eur Radiol. 2022;32(6):3672–82. https://doi.org/10.1007/s00330-021-08476-7.
Mo Z, Xin J, Chai R, Woo PYM, Chan DTM, Wang J. Epidemiological characteristics and genetic alterations in adult diffuse glioma in East Asian populations. Cancer biology Med. 2022;19(10):1440–59. https://doi.org/10.20892/j.issn.2095-3941.2022.0418.
Liu Y, Chen H, Li G, Zhang J, Yao K, Wu C, Li S, Qiu X. Radiotherapy delays malignant transformation and prolongs survival in patients with IDH-mutant gliomas. Cancer biology Med. 2022;19(10):1477–86. https://doi.org/10.20892/j.issn.2095-3941.2022.0472.
Park YW, Kim S, Park CJ, Ahn SS, Han K, Kang SG, Chang JH, Kim SH, Lee SK. Adding radiomics to the 2021 WHO updates may improve prognostic prediction for current IDH-wildtype histological lower-grade gliomas with known EGFR amplification and TERT promoter mutation status. Eur Radiol. 2022;32(12):8089–98. https://doi.org/10.1007/s00330-022-08941-x.
Ma C, Yao Z, Zhang Q, Zou X. Math Biosci engineering: MBE. 2020;18(1):727–44. https://doi.org/10.3934/mbe.2021039. Quantitative integration of radiomic and genomic data improves survival prediction of low-grade glioma patients.
Zhang H, Fan X, Zhang J, Wei Z, Feng W, Hu Y, Ni J, Yao F, Zhou G, Wan C, Zhang X, Wang J, Liu Y, You Y, Yu Y. Deep-learning and conventional radiomics to predict IDH genotyping status based on magnetic resonance imaging data in adult diffuse glioma. Front Oncol. 2023;13:1143688. https://doi.org/10.3389/fonc.2023.1143688.
Johnson DR, Sawyer AM, Meyers CA, O'Neill BP, Wefel JS. Early measures of cognitive function predict survival in patients with newly diagnosed glioblastoma. Neurooncology. 2012;14(6):808–16. https://doi.org/10.1093/neuonc/nos082.
Yang J, Lian JW, Chin YH, Wang L, Lian A, Murphy GF, Zhou L. Assessing the Prognostic Significance of Tumor-Infiltrating Lymphocytes in Patients With Melanoma Using Pathologic Features Identified by Natural Language Processing. JAMA Netw open. 2021;4(9):e2126337. https://doi.org/10.1001/jamanetworkopen.2021.26337.
Cepeda S, García-García S, Arrese I, Velasco-Casares M, Sarabia R. Relationship between the overall survival in glioblastomas and the radiomic features of intraoperative ultrasound: a feasibility study. J Ultrasound. 2022;25(1):121–8. https://doi.org/10.1007/s40477-021-00569-9.
d'Ocagne M. (1909). Calcul graphique et nomographie. Nature, 79(2045), 279–279. https://doi.org/10.1038/079279a0.

No competing interests reported.

20240711SupplementaryMaterials.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

From Imaging to Prognosis: Crafting Clinical Nomograms Based on a Multi-Sequence MRI Radiomics Model for Non- Invasive Glioma Survival Prediction

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

1. Introduction

2. Materials and Methods

2.1 Patient enrolment

2.2 MRI data acquisition and preprocessing

2.4 VOI segmentation and feature extraction

2.5 Preliminary Screening of Radiomics Features

2.6 Machine-learning-based prognostic prediction Feature selection and model construction

2.7 Evaluation

2.9 Statistical Analysis and data availability

3. Results

3.1 Comparative Evaluation of CoxFS, CoxLASSO, and RSFVH Feature Selection Methods

3.2 Radiomics Model Demonstrates Superior Prognostic Efficacy for GBM Compared to non-GBM

3.3 Evaluation of Models for Six Individual MRI Sequences

3.4 Evaluation of Models for Two Multi-MRI Sequences

3.5 Leveraging the Optimal Combo-4 Model for Glioma Survival Prediction

3.6 Risk Stratification Proficiency of RadSurv

3.7 Model Validation on Cox Regression Analysis

3.8 Nomogram Development and Validation

4. Discussion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1