Deep learning based algorithm for postoperative glioblastoma MRI segmentation: a promising new tool for tumor burden assessment

doi:10.21203/rs.3.rs-2729289/v1

Objective: Clinical and surgical decisions for glioblastoma patients depend on a tumor imaging-based evaluation. Artificial Intelligence (AI) can be applied to Magnetic Resonance Imaging (MRI) assessment to support clinical practice. In a real-world context, the current obstacles for AI are low-quality imaging and post-operative reliability. The aim of this study is to train an automatic algorithm for glioblastoma segmentation on a clinical MRI dataset and to obtain reliable results both pre- and post-operatively.

Methods: The dataset used for this study comprises 237 (71 preoperative and 166 postoperative) MRIs from 71 patients affected by an histologically confirmed Grade IV Glioma. The implemented U-Net architecture was trained by transfer learning to segment the resection cavity (RC), the enhancing tumor volume (ET) and the whole tumor volume (WT) in post-operative images. The learning phase was carried out first on BraTS2021 dataset for preoperative segmentation. Performance is evaluated using DICE score (DS) and Hausdroff 95%.

Results: The overall mean DS are 91.09 (± 0.60) and 72.31 (± 2.88), respectively obtained in pre-operative and post-operative scenarios. The overall mean Hausdroff 95% is 8.35 (± 1.12) for preoperative segmentation and 23.43 (± 7.24) for the postoperative assessment. Remarkably, the RC segmentation obtained a mean DS of 63.52 (± 8.90) in postoperative MRIs.

Conclusions: The performances achieved by the algorithm are consistent with previous literature for both pre-operative and post-operative glioblastoma’s MRI evaluation. Through the proposed algorithm, it is possible to reduce the impact of low-quality images and missing sequences.

glioma

glioblastoma

magnetic resonance imaging

deep learning

machine learning

segmentation

Glioblastoma is the most common adult malignant primary brain tumor (Bianconi et al., 2022; Venur et al., 2015). Despite the advances in therapeutic options and management, survival of patients with glioblastoma remains around 15-18 months, with a 5-year overall survival of approximately 5% (Bruno et al., 2022; Davis et al., 2012; Saaid et al., 2022). In order to improve the prognosis of the disease, research attempts are focused on new tools to individualize the therapeutic approach according to clinical, radiological, and molecular characteristics. Magnetic Resonance Imaging (MRI) is an essential tool for glioblastoma evaluation, providing necessary information to set up the best therapeutic strategy, both in preoperative decisions and in postoperative management, for each individual patient. Furthermore, MRI segmentation allows for volumetric assessment of different tumor components and, thus, precise surgical planning (Cordova, Schreibmann, Hadjipanayis, Guo, Hui-Kuo, et al., 2014; Zeppa et al., 2020). The current standard of MRI segmentation for tumoral lesions relies on manual measurements; however, this method is not only time-consuming but also not reproducible due to inter-operator variability (Berntsen et al., 2020; Cordova, Schreibmann, Hadjipanayis, Guo, Shu, et al., 2014; Gooya et al., 2012; Huber et al., 2017; Kubben et al., 2010; Menze et al., 2015a; Porz et al., 2014). For this reason, semi-automatic and completely automatic segmentation algorithms have been developed and evaluated in previous studies (Ghaffari et al., 2020; Mittal et al., 2022). With the recent development of Deep Convolutional Neural Networks (DCNNs) such as U-Net, now widely used for segmentation of medical imaging, the level of accuracy has increased ^{(Weng & Zhu, 2015)}.

Growing efforts to apply Artificial Intelligence (AI) to medical imaging analysis has resulted in the availability of larger datasets and improvements to software performances(Kaynak, 2021). For research purposes, MRIs have been collected and benchmarked in large datasets, such as the Brain Tumor Image Segmentation (BraTS) dataset (Ghaffari et al., 2020; Menze et al., 2015b).

Despite the good results obtained in the evaluation of preoperative images (Cordova, Schreibmann, Hadjipanayis, Guo, Shu, et al., 2014; Eijgelaar et al., 2020; Fyllingen et al., 2016; Kommers et al., 2021), AI MRI segmentation is limited in its postoperative evaluation and external validity. Limitations in postoperative MRI evaluation are partly due to artefacts, caused by blood and air in the resection cavity, and logistical issues in collecting data regularly from the same patient during follow-up. Moreover, the results obtained from most of the algorithms are not easily reproducible in the clinical context since they are frequently trained on curated and standardized datasets that do not include suboptimal quality images.

Obtaining a reliable system for postoperative automatic segmentation of MRI would bring many benefits to clinical practice. First, automatic segmentation would guarantee a fast and objective evaluation of MRI. This would be useful in optimizing and personalizing administration of treatment specific to unique features present in each case. Additionally, a well-trained algorithm may be able to perceive more than human experts can, possibly obtaining further information about the likely disease progression. Finally, it is important to train the algorithm to be considerate of common clinical practice issues, such as the heterogeneity of data and the frequent lack of MRI sequences.

DATASET

The Neurosurgery unit of the hospital Molinette (AOU Città della Salute e della Scienza di Torino) developed a dataset comprising 71 patients who underwent surgery at the institute, with histologically confirmed Glioblastoma Grade IV. All MRI scans acquired in situ were accessible by default on the hospital SYNAPSE®Mobility PACS system. Scans acquired in a different center were routinely uploaded onto the BRAINLAB©neuronavigation system of the neurosurgery department. At the time of hospitalization, written consent for personal, biological, and radiological data processing for scientific purposes was explicitly asked and registered on the InterSystem TrakCare©information system. The present study was approved by the local Institutional Review Board (n. 00162/2022). The following exclusion criteria were applied: underage subject (<18 years old), absence of T1 contrast enhancing (T1ce) or Fluid Attenuated Inversion Recovery (FLAIR) MRI sequences, post-operative complications (eg. hemorrhage or abscess) which could possibly invalidate the segmentation, or absence of histological confirmation. For each patient, one or more scans were available, including preoperative and postoperative images, with acquisition-time ranging from immediately after the surgery (max 48h) to 12 months later. For each MRI scan, a volumetric T1ce sequence and a FLAIR sequence were added to the radiological database.

The data were anonymized before processing, as instructed by the EU General Data Protection Regulation, using the specific function available in the HOROS©DICOM image viewer.

SEMI-AUTOMATIC SEGMENTATION

Both preoperative and postoperative segmentations were performed semi-automatically through the SmartBrush feature of the Cranial Planning workflow inside the BRAINLAB© neuronavigation system (Build 3.3.1.404). The volumetric representation was reconstructed by the software by combining the semi-automatic segmentations in the axial, coronal, and sagittal planes. The axial view was then extracted and, if necessary, manually adjusted. The process was carried out individually by 4 neurosurgeons, and 1 medical student. All the segmentations were revised and confirmed by a senior neurosurgeon with 25-year expertise in neurooncology and a neuroradiologist.

The segmented classes consisted of necrosis, core tumor (necrosis plus contrast-enhancing), and whole tumor (WT) (core tumor plus FLAIR hyperintensity) for the preoperative images; resection cavity (RC), enhancing tumor (ET), gross tumor volume (GTV) and WT (ET+ edema) were the classes for the postoperative images. For both preoperative and postoperative cases, whole tumor segmentation was performed on the FLAIR modality, whereas the remaining two classes on the T1ce one.

HARMONIZATION

To train the model on the BraTS dataset, MRIs have been harmonized accordingly.

At first, co-registration to the SRI-24 template with uniform isotropic resolution (1mm³) was performed (Rohlfing et al., 2010). The whole pipeline is available through the Cancer Imaging Phenomics Toolkit (Davatzikos et al., 2018; Pati et al., 2020), but it was only followed for the SRI-24 atlas co-registration portion. Skull-stripping was performed through SynthStrip (Hoopes et al., 2022).

The CaPTk pipeline required the presence of four scan modalities - FLAIR, T1, T1ce, and T2 – but only T2-FLAIR and T1ce were always present in the dataset. Hence, the burned-in segmentations were used as T1 (burned-in T1ce) and T2 (burned-in FLAIR). SynthStrip was then applied to extract the brain mask from the original end-of-pipeline T1ce scan, with the resulting mask adopted for all remaining scans (Figure 1).

Once both raw and burned-in data were registered to the SRI-24 atlas and skull-stripped, all classes were extracted by retrieving high-intensity regions. The complete segmentation was built by joining the necrosis/RC (respectively for pre/post-operative scans), the ET, and the peripheral edema.

PROPOSED ARCHITECTURE

The selected architecture was the winning contribution to the BraTS 2021 challenge, a 3D U-Net derived from the nnU-Net framework.(Futrega et al., 2021). Therefore, following the official NVIDIA GitHub repository, the nnU-Net variant, the preprocessing and the training schedule were implemented. Pre-training on preoperative BraTS dataset was performed for 150 epochs while all other relevant hyperparameters were kept as presented in the work by Furtega et al.(Futrega et al., 2021) The network was trained on a High Performance Computing (HPC) server using one 32G V100 nVIDIA GPUs. One of the challenges do be the tackled was the varying number of sequences available in clinical studies. Indeed, T1 and T2 sequences were frequently absent in the dataset, whereas existing architectures for pre-operative sequences are trained to produce the segmentation based on four modalities. This issue can be tackled in several ways: on one hand, one could simply remove sequences that are not commonly available in all studies to train the post-operative segmentation network.

On the other hand, one could seek to artificially synthesize the missing modalities using image modality translation (IMT) techniques (Figure 2).

These artificial sequences, although in a way “hallucinated” by the network, could be used in place of missing sequences, thus allowing the network to be trained on studies with a variable number of sequences. In this study, both strategies were compared. As IMT technique, the modified U-Net structure originally proposed by Osman and Tamam ^{(Osman & Tamam, 2022b)} was implemented to artificially synthesize these sequences, stacking T1ce and T2-FLAIR, hence exploiting information from both of them, instead of simply using a single sequence as input. The network was trained for 50 epochs with mixed precision on the BraTS 2021 dataset. The underlying U-Net required 2D inputs, therefore for each iteration a 3D scan was selected and a random selection of 64 2D slices, possibly flipped along the two dimensions, was extracted and fed as a batch. The network was trained with the default Adam optimizer with a learning rate of 0.0003, which decays following a cosine schedule. Output images were shaped 224×224 so post-processing was applied to resize, re-orient and pad them in their BraTS form. Gaussian sharpening was also applied as further data augmentation. The synthesis was carried out using the T1ce MRI acquisition to avoid noisy outputs eventually generated by the presence of non-volumetric FLAIR scans.

The network proposed by Futrega et al. (Futrega et al., 2021)and pre-trained on preoperative brain tumor segmentation, was fine-tuned on the Molinette post-operative dataset to reap the benefit of Transfer Learning (TL) (Hadjiiski et al., 2022). Several TL –strategies have been investigated to experimentally determine the best balance between expressivity and overfitting. All configurations with frozen layers in depth levels 1–7 have been studied, but no performance improvement or relevant advancement were observed. As a result, fine-tuning, consisting of tuning all weights from the pre-trained model, was chosen for the final experiments.

Training was performed for 200 epochs with a learning rate of 0.0007, which decays following a cosine schedule having 0.25 as cosine cycles parameter. The other relevant hyperparameters were kept the same as in the preoperative cases. A more aggressive data augmentation strategy was adopted by increasing the probability of applying a given transformation from 0.15 to 0.5: as suggested by Zhang et al., a more aggressive data augmentation might help in improving the model´s generalizability when trained on a small dataset(Zhang et al., 2020).

PERFORMANCE ANALYSIS

Cross-Validation

A 5-fold cross-validation was performed on the Molinette dataset with the purpose of maximizing the use of available data, while providing a more reliable estimation(Hadjiiski et al., 2022). The dataset was split at the patient level to ensure that different folds are statistical independent, i.e., the network is not trained and tested on scans from the same patient. Thus, it was possible to estimate how well the model generalizes on never-before-seen individuals.

Non-volumetric scans were included in order to have a dataset representative of common clinical scenarios. Nevertheless, patients with non-volumetric scans were only included in the training set and not in the validation set, since nowadays non-volumetric scans are less frequent in clinical practice, and thus it was better to test only on standard scans without compromising the validation of the algorithm.

Post-Processing

Post-processing was applied in order to bring back the three overlapping regions to the default ones of edema, ET, and RC. Specifically, if the probability of a voxel being classified as either GTV or edema was ≤0.45, it was classified as background. Else, if the probability of being classified as GTV was ≤0.4, then the voxel was classified as edema. If the probability of being a GTV voxel was >0.4 and the same held for the probability of being classified as RC, then the voxel was identified as such (or simply ET if the latter was not true). Moreover, any connected component identified as RC smaller than 16 voxels with an overall probability smaller than 0.9 was ignored and classified as ET instead (Figure 3).

These values were determined via a grid search on the 5 folds, starting from the parameters proposed by Futrega et al.(Futrega et al., 2021) for the preoperative case.

The dataset comprised 237 MRIs (71 preoperative and 166 postoperative) from 71 patients. The characteristics and number of those MRIs are summed up in Table 1.

Number of MRI scans	Number of patients	MRI scan type	Total
1	22	NC: 5 \| NE: 7 \| C: 10	22 (NV: 4)
2	25	NC: 3 \| NE: 14 \| C: 33	50 (NV: 9)
3	9	NC: 8 \| NE: 8 \| C: 11	27 (NV: 12)
4+	15	NC: 12 \| NE: 22 \| C: 33	67 (NV: 21)
Total	71	NC: 28 \| NE: 51 \| C: 87	166 (NV: 46)

Table 1: Quantitative description of the post operative dataset from Molinette Hospital (NC: no cavity, NE: no enhancing, C: complete, NV: non-volumetric).

PREOPERATIVE RESULTS

Evaluation was performed on two different configurations of available scans: the “complete” one, i.e. the one comprising all four modalities (FLAIR, T1, T1ce and T2), and the “most-informative subset” one, i.e. the one comprising just FLAIR and T1ce modalities. The complete configuration included as T1 and T2 the ones artificially synthesized through the 2D U-Net IMT method described above.

The overall mean DS and Hausdroff 95 distance are 91.09 (± 0.60) and 8.35 (± 1.12) respectively for the “complete” subset and 90.77 ± 0.67 and 8.35 ± 1.12 for the “most-informative” one. (Figure 4)

Results show that, even if optimal performance is obtained on the “complete” configuration, solid outcomes are also achieved with the two most informative MRI scans.

Results obtained during a 5-fold cross-validation of the BraTS 2021 dataset are in line with those presented by Futrega et al. (Dice Score (DS): 91.63). The result obtained (DS: 91.09) confirms the foundation of such an nnU-Net implementation for preoperative brain tumor segmentation in MRI.

POSTOPERATIVE RESULTS

As for preoperative scans, evaluation was performed for “complete” subset, and the “most-informative subset” one. The overall mean DICE score was 72.31 (± 2.88), while Hausdroff 95 distance was 23.43 (± 7.24) for the postoperative “most informative subset” assessment, as shown in figure 5. Similar results are obtained with the “complete” subset (DS: 72.31 ± 2.88, HD95: 23.85 ± 7.20) (Figure5).

Remarkably, the RC segmentation obtained a mean DS of 63.52 (± 8.90) in postoperative MRIs. (Table 2)

Model	Available modalities	All (Dice Score)	GTV (Dice Score)	Cavity (Dice Score)	Mean (Dice score)	All (Hausdroff95 distance)	GTV (Hausdroff95 distance)	Cavity (Hausdroff95 distance)	Mean (Hausdroff95 distance)
Fold 0	FLAIR, T1ce	74.45	74.52	69.77	72.91	15.20	9.96	8.78	11.31
Fold 0	FLAIR, T1, T1ce, T2	73.07	72.70	71.23	72.33	14.77	10.70	12.07	12.51
Fold 1	FLAIR, T1ce	77.80	74.89	54.78	69.16	8.09	16.36	59.18	27.88
Fold 1	FLAIR, T1, T1ce, T2	76.62	82.39	50.05	69.69	7.99	15.87	57.70	27.19
Fold 2	FLAIR, T1ce	76.10	71.41	55.93	67.81	11.40	11.60	75.23	32.74
Fold 2	FLAIR, T1, T1ce, T2	75.07	75.48	55.85	68.80	13.71	12.55	75.98	34.08
Fold 3	FLAIR, T1ce	77.16	81.23	71.82	76.74	7.83	7.03	47.00	20.62
Fold 3	FLAIR, T1, T1ce, T2	76.86	81.38	71.96	76.73	7.83	6.59	46.30	20.24
Fold 4	FLAIR, T1ce	77.02	80.43	69.31	75.59	9.34	6.56	57.94	24.61
Fold 4	FLAIR, T1, T1ce, T2	73.93	79.58	68.50	74.00	10.00	6.96	58.70	25.22
Mean+-STD	FLAIR, T1ce	76.51 ± 1.16	76.50 ± 3.75	64.32 ± 7.38	72.44 ± 3.49	10.37 ± 2.72	10.30 ± 3.56	49.63 ± 22.32	23.43 ± 7.24
Mean+-STD	FLAIR, T1, T1ce, T2	75.11 ± 1.48	78.31 ± 3.67	63.52 ± 8.90	72.31 ± 2.88	10.86 ± 2.88	10.53 ± 3.49	50.15 ± 21.27	23.85 ± 7.20

Table 2: Experimental Dice scores and Hausdroff95 distances obtained during 5-fold cross-validation on the postoperative Molinette Hospital dataset for the two available modalities configurations.

Again, the lack of T1 and T2 has not been proven to significantly impact the results achieved. The addition of synthetic T1 and T2 scans did not improve the overall segmentation, as suggested by previous literature.

Graphical exemples of automatic segmentation with difference map from manual segmentation, both for preoperative and postoperative MRI from Molinette database, are shown respectively in Figure 6A and 6B.

As previously stated, currently the gold standard for volumetric assessment is manual segmentation. Numerous efforts have been taken to improve volumetric assessment and segmentation of the tumoral lesion in the preoperative context (Baid et al., 2021; Eijgelaar et al., 2020; Fyllingen et al., 2016; Kommers et al., 2021).

Available automatic algorithms were developed mainly for preoperative images; this results in low reliability for postoperative assessment (Zeppa et al., 2020).

The practical reason behind this is the intrinsic difficulty in postoperative MRI segmentation (Baid et al., 2021; Cordova, Schreibmann, Hadjipanayis, Guo, Shu, et al., 2014). In fact, the RC is frequently a source of artefacts in the MRI because of blood residuals and air bubbles (Ermiş et al., 2020b; Visser et al., 2019). In addition to this, brain anatomy may be partly altered as a consequence to the surgical act, the postsurgical edema and the tumor itself (Visser et al., 2019). These problems lower the accuracy of available algorithms in obtaining postoperative evaluation of MRI, in addition to logistical issues concerning regular post-surgical follow-up (Chang et al., 2019). Nevertheless, some studies recently reported good accuracy in postoperative segmentation of MRI, though it is still far from the level of accuracy achieved in preoperative evaluation (Chang et al., 2019; Cordova, Schreibmann, Hadjipanayis, Guo, Shu, et al., 2014).

Another limit of the available algorithms is that they are often trained on cured and standardized datasets that do not include low-quality images. Though this selection bias makes the training process easier, it is not as easily transferable to real-world clinical practice. In fact, suboptimal quality of data is very common in clinical practice, including non-volumetric scans, missing sequences, and artifacts (Ermiş et al., 2020a).

In this study, we aimed to train an AI algorithm for the postoperative MRI evaluation of glioblastoma in order to prospectively introduce this tool in clinical practice as support for the decision-making process. For this reason, the MRI database used for the training is representative of the real-world clinical scenario, frequently including heterogeneous and incomplete data. We did not apply restrictive inclusion criteria concerning the quality of the available data in order not to affect the results with selection bias.

Low-quality images were also included, concerning especially non-volumetric MRIs. The presence of non-volumetric images is related to old acquisition protocols, but their presence in the clinical scenario is still relevant, accounting for almost 25% of the MRIs collected in the Molinette database. As a consequence, this data was not excluded from the study as it would limit prospective application of the algorithm in clinical practice.

Moreover, postoperative images have different acquisition times given the time-course of the disease and the treatment schedule. This means that the postoperative MRI database contains images from different points in time: immediate postoperative, before and after adjuvant treatment, and regular follow-up. Herein, the algorithm is exposed to different biological entities such as post-surgical residual, RC, progressively growing lesion, and edema.

Herein, the results achieved are similar to the ones reported in other studies, considering both preoperative (mean DS: 91.09 ± 0.60) and postoperative (mean DS: 72.31 ± 2.88) evaluation. From the results obtained, it is evident that the accuracy in the postoperative setting is still far away from that in the preoperative scenario. This contrast in accuracy is especially remarkable for the RC segmentation, with a mean DS of 63.52 ± 8.90. This element causes both cases of hyper-segmentation, including adjacent regions, and sub-segmentation, excluding some parts of the cavity. Nevertheless, the evaluation of the RC is complex with less accurate results even for expert human operators performing manual segmentation.

The level of accuracy reached in this study was improved by the application of various informatic strategies. In particular, TL, data augmentation, cross-validation and an ensemble of models aggregated through the STAPLE algorithm compensate for the limited amount of data. ²². Another challenge for applying automatic segmentation in clinical practice is the variable number of sequences available. IMT is a technique that takes information from existing sequences to create the missing ones, but it is still at an experimental level. In this study, IMT architecture from Osman et al. ^{(Osman & Tamam, 2022a)} was applied to T1ce sequences to create T1 and T2 whenever they were not available in the Molinette database. In preoperative segmentation, even if these sequences are nonessential, they improve the performance of the algorithm. In our study, we did not observe any benefits associated with IMT, unlike suggested by previous literature.(Yang et al., 2020) However, it is possible that with larger or more diverse datasets the quality of the synthesized images could be improved, especially in the post-operative setting.

LIMITS OF THE STUDY

Institutional studies with private datasets are essential to scientific and informatic research, but they have some limitations (Petrick et al., 2013). Literature reports that models developed and tested with data from one collection hardly achieve similar results when applied to data from a different institute (Wei et al., 2019). It is therefore advisable to corroborate the results from this study with multi-institutional data, consequently increasing the level of reliability.

In addition, several studies highlight that reference standards based on the expertise of radiologists are not completely objective (Revesz et al., 1983). It is reported that the number of operators performing the segmentation should be at least three (Petrick et al., 2013), while, in this study, the manual segmentation was performed by four neurosurgeon, one medical student and revised by a senior neurosurgeon and a neuroradiologist in order to overcome interobserver variability.

A further limitation in the proposed work is the final post-processing pipeline proposed to bring back labels to tumoral segments of the postoperative evaluation (edema, enhancing tumor and resection cavity). Even if the parameters are obtained by averaging grid-search outputs, the limited amount of data decreases the reliability of these values. Indeed, since the network is less confident in its predictions than in the preoperative cases, it is plausible that such prediction confidence might not be above the chosen threshold for some MRI scans, leading to imprecise segmentations.

As the training phase influences the outcomes of the algorithm, quality assessment of MRIs used in this step would be helpful. Moreover, results would be more accurate if the T2-FLAIR sequence was always volumetric, however, the purpose of this work was to avoid selection bias of data to get an algorithm reliable in the clinical practice. For this reason, the possible improvement relies on more accurate protocols of MRI acquisition in common clinical practice and not on image selection for research studies.

FUTURE PERSPECTIVES

Due to the benefits granted by informatic tools and strategies, our results are in line with the existing literature on this topic. Different from previous studies, this work is not biased by restrictive inclusion/exclusion criteria for MRI scans. Therefore, we present this work as a starting point to apply AI to clinical practice for glioblastoma with remarkable reliability both in the preoperative and postoperative context.

Future studies should involve multiple institutions, allowing for an increase in the sample size of the database overall and of glioblastoma postoperative MRIs acquired from different protocols and machines. Moreover, experimental techniques such as IMT could be refined, adding greater support to the algorithm. The elimination of non-volumetric scans and low-quality imaging from clinical practice would be essential not only for research purposes but also for future clinical application of the AI technologies. All of these initiatives may improve the AI algorithm performance and lead to clinically reliable use of AI in glioblastoma evaluation.

Finally, working with AI requires simultaneous specialized technical competences and a comprehensive view of the clinical scenario. Thus, it is advisable to face the current limitations of biological, clinical, logistical, and technical issues within the analysis from a multidisciplinary point of view. This outlook highlights the importance of clear communication between the neurosurgical team and the engineers in searching for appropriate solutions.

This study sought to create a reliable tool for both preoperative and postoperative automatic MRI segmentation of glioblastoma. The algorithm obtained still has some limitations, but the results of the study are in line with the existing literature. Moreover, the authors chose to train the algorithm to be reliable in clinical practice, especially in cases of missing sequences or low-quality images. Some strategies have been proposed in this work to overcome these limitations, with promising results. In the future, the remaining challenges ahead may be faced by increasing the dataset size and implementing innovative technical strategies.

Acknowledgments: Computational resources were provided by HPC@POLITO, a project of Academic Computing within the Department of Control and Computer Engineering at the Politecnico diTorino (http://www.hpc.polito.it).

Conflicts of Interest: The authors declare no conflict of interest.

Venur VA, Peereboom DM, Ahluwalia MS. Current Medical Treatment of Glioblastoma. In: ; 2015:103-115. doi:10.1007/978-3-319-12048-5_7
Bianconi A, Aruta G, Rizzo F, et al. Systematic Review on Tumor Microenvironment in Glial Neoplasm: From Understanding Pathogenesis to Future Therapeutic Perspectives. Int J Mol Sci. 2022;23(8):4166. doi:10.3390/ijms23084166
Davis FG, Dolecek TA, McCarthy BJ, Villano JL. Toward determining the lifetime occurrence of metastatic brain tumors estimated from 2007 United States cancer incidence data. Neuro Oncol. 2012;14(9):1171-1177. doi:10.1093/neuonc/nos152
Bruno F, Pellerino A, Pronello E, et al. Elderly Gliobastoma Patients: The Impact of Surgery and Adjuvant Treatments on Survival: A Single Institution Experience. Brain Sci. 2022;12(5). doi:10.3390/BRAINSCI12050632
Saaid A, Monticelli M, Ricci AA, et al. Prognostic Analysis of the IDH1 G105G (rs11554137) SNP in IDH-Wildtype Glioblastoma. Genes (Basel). 2022;13(8). doi:10.3390/GENES13081439
Zeppa P, Neitzert L, Mammi M, et al. How reliable are volumetric techniques for high-grade gliomas? A comparison study of different available tools. Neurosurgery. 2020;87(6):E672-E679. doi:10.1093/neuros/nyaa282
Cordova JS, Schreibmann E, Hadjipanayis CG, et al. Quantitative tumor segmentation for evaluation of extent of glioblastoma resection to facilitate multisite clinical trials. Transl Oncol. 2014;7(1):40-47. doi:10.1593/tlo.13835
Porz N, Bauer S, Pica A, et al. Multi-Modal Glioblastoma Segmentation: Man versus Machine. PLoS One. 2014;9(5):e96873. doi:10.1371/journal.pone.0096873
Menze BH, Jakab A, Bauer S, et al. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Trans Med Imaging. 2015;34(10):1993-2024. doi:10.1109/TMI.2014.2377694
Kubben PL, Postma AA, Kessels AGH, van Overbeeke JJ, van Santbrink H. Intraobserver and Interobserver Agreement in Volumetric Assessment of Glioblastoma Multiforme Resection. Neurosurgery. 2010;67(5):1329-1334. doi:10.1227/NEU.0b013e3181efbb08
Cordova JS, Schreibmann E, Hadjipanayis CG, et al. Quantitative Tumor Segmentation for Evaluation of Extent of Glioblastoma Resection to Facilitate Multisite Clinical Trials. Transl Oncol. 2014;7(1):40-W5. doi:10.1593/tlo.13835
Gooya A, Pohl KM, Bilello M, et al. GLISTR: Glioma Image Segmentation and Registration. IEEE Trans Med Imaging. 2012;31(10):1941-1954. doi:10.1109/TMI.2012.2210558
Huber T, Alber G, Bette S, et al. Reliability of Semi-Automated Segmentations in Glioblastoma. Clin Neuroradiol. 2017;27(2):153-161. doi:10.1007/s00062-015-0471-2
Berntsen EM, Stensjøen AL, Langlo MS, et al. Volumetric segmentation of glioblastoma progression compared to bidimensional products and clinical radiological reports. Acta Neurochir (Wien). 2020;162(2):379-387. doi:10.1007/s00701-019-04110-0
Ghaffari M, Sowmya A, Oliver R. Automated Brain Tumor Segmentation Using Multimodal Brain Scans: A Survey Based on Models Submitted to the BraTS 2012–2018 Challenges. IEEE Rev Biomed Eng. 2020;13:156-168. doi:10.1109/RBME.2019.2946868
Mittal H, Pandey AC, Saraswat M, Kumar S, Pal R, Modwel G. A comprehensive survey of image segmentation: clustering methods, performance parameters, and benchmark datasets. Multimed Tools Appl. 2022;81(24):35001-35026. doi:10.1007/s11042-021-10594-9
Weng W, Zhu X. U-Net: Convolutional Networks for Biomedical Image Segmentation. IEEE Access. 2015;9:16591-16603. doi:10.48550/arxiv.1505.04597
Kaynak O. The golden age of Artificial Intelligence. Discover Artificial Intelligence. 2021;1(1). doi:10.1007/s44163-021-00009-x
Menze BH, Jakab A, Bauer S, et al. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Trans Med Imaging. 2015;34(10):1993-2024. doi:10.1109/TMI.2014.2377694
Fyllingen EH, Stensjøen AL, Berntsen EM, Solheim O, Reinertsen I. Glioblastoma segmentation: Comparison of three different software packages. PLoS One. 2016;11(10). doi:10.1371/journal.pone.0164891
Eijgelaar RS, Visser M, Müller DMJ, et al. Robust deep learning–based segmentation of glioblastoma on routine clinical MRI scans using sparsified training. Radiol Artif Intell. 2020;2(5):1-9. doi:10.1148/ryai.2020190103
Kommers I, Bouget D, Pedersen A, et al. Glioblastoma Surgery Imaging - Reporting and Data System: 2standardized reporting of tumor volume, location, and resecta-3bility based on automated segmentations. Cancers (Basel). 2021;13. doi:10.3390/xxxxx
Rohlfing T, Zahr NM, Sullivan E v., Pfefferbaum A. The SRI24 multichannel atlas of normal adult human brain structure. Hum Brain Mapp. 2010;31(5):798-819. doi:10.1002/HBM.20906
Pati S, Singh A, Rathore S, et al. The Cancer Imaging Phenomics Toolkit (CaPTk): Technical Overview. In: ; 2020:380-394. doi:10.1007/978-3-030-46643-5_38
Davatzikos C, Rathore S, Bakas S, et al. Cancer imaging phenomics toolkit: quantitative imaging analytics for precision diagnostics and predictive modeling of clinical outcome. Journal of Medical Imaging. 2018;5(01):1. doi:10.1117/1.JMI.5.1.011018
Hoopes A, Mora JS, Dalca A v., Fischl B, Hoffmann M. SynthStrip: skull-stripping for any brain image. Neuroimage. 2022;260:119474. doi:10.1016/j.neuroimage.2022.119474
Futrega M, Milesi A, Marcinkiewicz M, Ribalta P. Optimized U-Net for Brain Tumor Segmentation. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2021;12963 LNCS:15-29. doi:10.48550/arxiv.2110.03352
Osman AFI, Tamam NM. Deep learning‐based convolutional neural network for intramodality brain MRI synthesis. J Appl Clin Med Phys. 2022;23(4). doi:10.1002/acm2.13530
Hadjiiski L, Cha K, Chan HP, et al. AAPM task group report 273: Recommendations on best practices for AI and machine learning for computer-aided diagnosis in medical imaging. Med Phys. Published online 2022. doi:10.1002/MP.16188
Zhang L, Wang X, Yang D, et al. Generalizing Deep Learning for Medical Image Segmentation to Unseen Domains via Deep Stacked Transformation. IEEE Trans Med Imaging. 2020;39(7):2531-2540. doi:10.1109/TMI.2020.2973595
Baid U, Ghodasara S, Mohan S, et al. The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification. Published online July 5, 2021. http://arxiv.org/abs/2107.02314
Ermiş E, Jungo A, Poel R, et al. Fully automated brain resection cavity delineation for radiation target volume definition in glioblastoma patients using deep learning. Radiation Oncology. 2020;15(1):100. doi:10.1186/s13014-020-01553-z
Visser M, Müller DMJ, van Duijn RJM, et al. Inter-rater agreement in glioma segmentations on longitudinal MRI. Neuroimage Clin. 2019;22:101727. doi:10.1016/j.nicl.2019.101727
Chang K, Beers AL, Bai HX, et al. Automatic assessment of glioma burden: A deep learning algorithm for fully automated volumetric and bidimensional measurement. Neuro Oncol. 2019;21(11):1412-1422. doi:10.1093/neuonc/noz106
Ermiş E, Jungo A, Poel R, et al. Fully automated brain resection cavity delineation for radiation target volume definition in glioblastoma patients using deep learning. Radiation Oncology. 2020;15(1). doi:10.1186/s13014-020-01553-z
Osman AFI, Tamam NM. Deep learning-based convolutional neural network for intramodality brain MRI synthesis. J Appl Clin Med Phys. 2022;23(4). doi:10.1002/ACM2.13530
Yang Q, Li N, Zhao Z, Fan X, Chang EIC, Xu Y. MRI Cross-Modality Image-to-Image Translation. Scientific Reports 2020 10:1. 2020;10(1):1-18. doi:10.1038/s41598-020-60520-6
Petrick N, Sahiner B, Armato SG, et al. Evaluation of computer-aided detection and diagnosis systems. Med Phys. 2013;40(8). doi:10.1118/1.4816310
Wei L, Roberts DW, Sanai N, Liu JTC. Visualization technologies for 5-ALA-based fluorescence-guided surgeries. J Neurooncol. 2019;141(3):495-505. doi:10.1007/S11060-018-03077-9/TABLES/2
Revesz G, Kundel HL, Bonitatibus M. The effect of verification on the assessment of imaging techniques. Invest Radiol. 1983;18(2):194-198. doi:10.1097/00004424-198303000-00018

No competing interests reported.

Deep learning based algorithm for postoperative glioblastoma MRI segmentation: a promising new tool for tumor burden assessment

Status:

Version 1

Abstract

Figures

Introduction

Methods

Results

Discussion

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1