Ultrasound-based deep learning radiomics nomogram for comprehensive prediction of tumor, axillary lymph node status and prognosis of breast cancer patients after neoadjuvant chemotherapy

doi:10.21203/rs.3.rs-3787783/v1

Download PDF

Research Article

Ultrasound-based deep learning radiomics nomogram for comprehensive prediction of tumor, axillary lymph node status and prognosis of breast cancer patients after neoadjuvant chemotherapy

https://doi.org/10.21203/rs.3.rs-3787783/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Objectives Our study aims to explore the feasibility of the deep learning radiomics nomogram (DLRN) for predicting the status of tumors and axillary lymph node metastasis (ALNM) after neoadjuvant chemotherapy (NAC) in breast cancer patients, and employ a Cox regression model for survival analysis to validate the effectiveness of the fusion algorithm.

Methods A total of 243 patients who underwent NAC were retrospectively included between October 2014 and July 2022. The DLRN integrated clinical characteristics as well as radiomics and deep transfer learning features extracted from ultrasound images. Evaluating the diagnostic performance of DLRN by constructing ROC curve , and assessing the clinical usefulness of models by using decision curve analysis (DCA). A survival model was developed to validate the effectiveness of the fusion algorithm.

Results In the training cohort, the DLRN yielded an area under the receiver operating characteristic curve (AUC) values of 0.984 and 0.985 in the tumor and LNM, while 0.892 and 0.870, respectively, in the test cohort. The consistency index (C-index) of the nomogram was 0.761 and 0.731, respectively, in the training and test cohort.The Kaplan-Meier survival curves showed that patients in the high-risk group had significantly poorer overall survival than patients in the low-risk group (P<0.05).

Conclusion The US-based DLRN model could hold promise as a clinical guidance for predicting the status of tumor and LNM after NAC in breast cancer patients. This fusion model also can predict the prognosis of patients , which could help clinicians to make better clinical decisions.

Neoadjuvant chemotherapy

Breast cancer

Ultrasound

Deep learning

Fusion

Neoadjuvant chemotherapy (NAC) is the main treatment method for patients with locally advanced breast cancer(BC). Its main purpose is to reduce the tumor volume and downstage primary tumors and eliminate lymph node metastasis (LNM), so that inoperable breast cancer patients or patients who are not suitable for breast conservation have the opportunity to receive surgical treatment^{[1, 2]}. The responses of the breast tumor and axillary lymph node metastasis (ALNM) to NAC are highly correlated with subsequent surgical protocols, disease-free survival (DFS) of patients, and overall survival (OS) ^{[3, 4]}. A comprehensive evaluation of NAC can help clinicians adjust treatment methods, improve disease-free survival and overall survival of patients, and has important clinical significance.

Previous studies have mostly predicted the status of tumor and ALNM after NAC based on imaging features at a certain time point. However, the morphological features of tumors and ALNM can change during NAC, and longitudinal monitoring of their dynamic changes is crucial for predicting the status of tumor and ALNM. Breast radiography and magnetic resonance imaging (MRI) have been used to monitor tumor response to NAC. However, mammography has radiation, MRI is expensive and the examination time is long, making it difficult to perform multiple repeated examinations in the short term^[5–8]. In contrast, ultrasound(US) is widely used to evaluate the treatment response of breast cancer due to its convenience and

non-radiation characteristics^[9].However, traditional ultrasound examination methods are not accurate enough, and surgical pathology is still the gold standard^[10]. Therefore, there is an urgent need for broader, available, and cost-effective tools to predict NAC treatment responses.

Radiomics is an emerging image analysis method that has been widely used in non-invasive prediction of neoadjuvant chemoradiotherapy responses in various malignant tumors^[11]. Multiple studies have shown that radiomics features play an important role in evaluating NAC responses^[12]. However, it need to manually segment the tumor^[13]. Better than conventional radiomics, deep transfer learning technology as a promising tool can autonomously learn complex hierarchical structures of image features and automatically extract high-level abstract features in neural networks without the need for precise tumor segmentation designed by humans as input^{[2, 14, 15]}.

Previous studies mostly used radiomics or deep learning features to evaluate the tumor or LNM response to NAC based on ultrasonography images before and after treatment^[16]. Radiomics features and deep transfer learning features both have their own advantages and disadvantages. Therefore, if these features are fused, they may reflect the potential complementary information. We assume that fusing these features may increase valuable information. To our knowledge,there are few studies that combine deep transfer learning features, radiomic features and clinical characteristics to evaluate the response to NAC in breast cancer. Thereby, the purpose of this study is to explore the feasibility of fusion model deep learning radiomics nomogram(DLRN) in predictint the status of tumors and LNM after NAC in patients with breast cancer, and employ a Cox regression model for survival analysis to validate the effectiveness of the fusion algorithm and facilitate clinical use.

Patients

This study was approved by the Hospital Ethics Committee. As this study is a retrospective study, no informed consent form was signed.

The inclusion criteria were as follows: (1) breast cancer and ipsilateral lymph node metastasis confirmed by biopsy; (2) completed all courses of NAC and underwent surgery after NAC; (3) having standard ultrasound examination; (4) having complete baseline clinicopathologic data. The exclusion criteria were as follows: (1) multiple lesions on the same side or bilateral lesions; (2) having not completed the entire NAC process; (3) poor image quality; (4) distant metastasis; (5) have a history of other cancers or systemic chemotherapy. In this study, a total of 243 patients who met inclusion criteria were included between October 2014 and July 2022.We randomly selected 80% of the samples as the training cohort, and the remaining 20% as the test cohort. We performed 5-fold cross-validation on the training cohort and used the Grid-Search algorithm to find optimal hyperparameters and optimize the algorithm.

Chemotherapy regimens and pathologic analysis

In accordance with National Comprehensive Cancer Network (NCCN) criteria, all included patients received six or eight rounds of standard NAC based on their molecular subtypes. The chemotherapy regimens mainly included either taxane based or anthracycline and taxane based, and patients with human epidermal growth factor receptor 2 (HER2)–positive were prescribed trastuzumab. The final tumor and ALNM status was determined by surgical pathology. The chemotherapy regimens and detailed pathologic analysis are provided in the Additional file 1: Table S1.

Image Acquisition

Breast US examinations were performed at before, during, and after NAC, by one of six radiologists with over 10 years of experience in breast US. They each convert the collected breast ultrasound images into Medical Digital Imaging and Communication (DICOM) format. Examinations were mainly performed using four different ultrasound machines: SuperSonic Imagine (Aixplorer), Siemens (S2000), Philips(EPIQ 7), GE(LOGIQ E9) which equipped with linear-array transducer. A typical image representing the largest diameter of the lesion was selected. In this way, poor-quality ultrasound images such as those with motion artifacts or incomplete breast lesions sections can be excluded, and the influence of the doctor's subjective perception and objective ability on the data set during image collection can be effectively avoided.

Image Segmentation

It is necessary to define a region of interest (ROI) of the paired US images of the same lesion before conducting NAC,during NAC and after NAC to eliminate the interference caused by irrelevant data, such as text and icons. The ROI in the were depicted by two radiologists all with more than 5 years of experience. The ROI segmentation was performed using open-source software, ITK-SNAP (http://www.itksnap.org/pmwiki/pmwiki.php).

Deep learning analysis

To eliminate the influence of surrounding background noise, we first cropped the annotated ROI region and resized all images to 256×256 pixels. For training frames, we applied online data augmentation and normalization, which included random horizontal and vertical flipping. Finally, the augmented frames were center-cropped to 224 × 224 pixels and z-score normalized on the RGB channels.

For frame-level prediction, we trained three classical convolutional neural networks (ResNet50, ResNet101, and DenseNet201) to predict the probabilities of the frames. The frames were assigned with the corresponding annotated label, and the network was optimized using a mini-batch gradient descent method with softmax cross-entropy (CE) loss. To improve the generalization performance, we carefully set the learning rate and employed the cosine decay learning rate algorithm in this study. More detailed can be seen in the Supplementary material.

Radiomics analysis

Feature extraction and selection

To reduce the impact of voxel spacing variation, spatial normalization is often employed. In our experiment, We resampled all the data to a resolution of 1mm×1mm to handle these issues. The handcrafted features which extracted by pyradiomics(https://pyradiomics.readthedocs.io/en/latest/) can be categorized into three groups: (I) geometry, (II) intensity, and (III) texture. Texture features were extracted using various methods such as gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM).

We conducted Mann-Whitney U-test statistical test and feature screening for all radiomics features. Features with high repeatability were further evaluated using Pearson's rank correlation coefficient to identify correlations between features. To maintain the ability to depict features as completely as possible, a greedy recursive deletion strategy was employed for feature filtering.The training data was used to perform LASSO regression model for feature selection. The model removes irrelevant features by setting their coefficients to zero, and 10-fold cross-validation is used to determine the optimal regulation weight λ. The retained features with non-zero coefficients are used to construct the radiomics signature.

The selected final features were utilized to construct a risk model employing various machine learning models such as Logistic Regression (LR), Support Vector Machine (SVM), AdaBoost, Multi-Layer Perceptron (MLP), XGBoost, LightGBM, etc.

Development of the fusion model

Instead of selecting only the best model as in previous works, we performed frame fusion on the predictions of all models to leverage the abilities of all models. For deep learning, we used the fine-tuned ResNet50, ResNet101, and DenseNet201 models to predict a result and its corresponding probability for each frame. For radiomics, we fused the prediction results of SVM, AdaBoost, MLP, XGBoost, and LightGBM models.

We combined the likelihoods of all frames using two independent methods: the Frames Likelihood Histogram (FLH) pipeline and the Bag of Words (BoW) pipeline. To perform feature selection, we utilized the PLH and BoW pipelines to aggregate the Frames likelihoods and TF-IDF features, respectively. The LASSO reel was then applied to select the most significant features, which were used as input for machine learning models following a similar modeling process to the radiomics. A schematic diagram for the integrated DLRN model is shown in Fig. 1.

Model Evaluation

In this study, we have two tasks as follows:

1. pCR/LNM vs. non-pCR/non-LNM:

The diagnostic performance of models were evaluated by constructing ROC curves. Calibration curves were plotted to assess the calibration performance of the nomogram, and the Hosmer-Lemeshow goodness-of-fit test was used to evaluate its calibration ability. Furthermore, decision curve analysis (DCA) was utilized to assess the clinical usefulness of the predictive models.

2. Survival analysis:

To validate the effectiveness of the fusion algorithm and facilitate clinical use, we developed a survival model. We fused the clinical features with the two different parts (tumor and LNM) and two pipelines (deep learning and radiomics) to construct a signature using univariable and multivariable analysis with a p-value < 0.05. We then used a Linear Cox proportional hazards model with L2 regularization (c = 0.1) to predict the survival model and visualize it with a nomogram. To evaluate the results of the survival model, we used Kaplan-Meier analysis to determine whether each model stratified patients into clinically significant groups. To delineate group membership, we tested percentile thresholds, choosing the value that maximized the significance of the separation in the training set by the log-rank test. This was performed individually for RFS, where relevant. All p-values for Kaplan-Meier analysis were calculated by the multivariate log-rank test.

Statistical analysis

Descriptive statistical data is summarized as mean standard deviation (SD).

Continuous variable t-test or Mann Whitney U-test were used for inter group comparison, and chi square test or Fisher's exact test were used for categorical variables. To determine the clinical usefulness, Decision curve analysis (DCA) was conducted to evaluate the clinical net benefit. P < 0.05 indicating statistically significant differences. All statistical analyses were conducted using R 3.4.1 and SPSS 23.0 (IBM, Armonk, NY, USA) software.

Clinical characteristics

Figure 2 illustrates the flowchart of patient selection. From October 2014 to July 2022, a total of 243 patients were included in the final analysis. The clinical characteristics of all patients were listed in Table 1(a,b). For the two cohorts, there was no significant difference between pCR and non-pCR or LNM and non‐LNM groups in terms of age,clinical N stage,histological type (P > 0.05). MDBT, MDAT were significantly correlated with pCR, while MSDAT was significantly correlated with LMN(P < 0.05).Meanwhile,clinical T stage, molecular subtypes, anti_HER2_treatment, the statuses of estrogen receptor (ER), progesterone receptor (PR), Ki-67 and HER2 were significantly associated with pCR or LNM in the training cohort as shown in Table 1(a,b).

Multivariable logistic regression analysis showed that MDAT、Ki-67 index were significant predictors of pCR, while HER2 was a significant predictor of LNM status, and they were selected as the key predictors for clinical model construction

(Additional file 1: Table S2 and Figure S1).

Consistent with previous studies^[17], the performance of clinical models in predicting pCR or LNM status were unsatisfactory, with AUC values of 0.834 (95% confidence interval (CI): 0.771–0.897), 0.693 (95%CI: 0.625–0.761) for tumor and LNM in the training cohort, while 0.810( (95%CI: 0.668–0.952), 0.738(95%CI: 0.613–0.864) in the test cohort, respectively.

Interpretation of Radiomics

We extracted a total of 1561 handcrafted radiomic features from each modality.The distribution of each group of handcrafted features is shown in Additional file 1: Figure S2. Additionally, we performed visualization of the local features of a specific sample, as shown in Additional file 1: Figure S2. We performed feature selection using Lasso. Finally, features with nonzero coefficients were selected to establish the Rad-score using a logistic regression model with least absolute shrinkage and selection operator (LASSO). Detailed information could be seen in Supplementary materials.

Interpretation of Deep Transfer Learning(DTL)

Gradient-weighted Class Activation Mapping (Grad-CAM) for visualization was implied to distinguish salient visual features, highlighting the input features with the largest weight in deep learning classification. To validate deep learning features, heatmaps were generated using the Grad-CAM method(Fig. 3). For patients with non-pCR or LNM, the peritumor region was highlighted, while the intratumor region was not in ultrasound images. For patients with pCR or non-LNM, the intratumor region was highlighted, whereas the peritumor region was not in ultrasound images.

Development and performance of model

We validated the fusion model with clinical and ultrasound data. We merged all features using feature concatenation. Table 2 described the diagnostic indexes of the different models in the training cohort and test cohorts.

For tumor, compared with the Clinical, Radiomics and DTL model, the DLRN showed the favorable AUC in the training cohort (0.984; 95% CI 0.972–0.997) and test cohort (0.892; 95% CI 0.789–0.996)(Fig. 4).In the training cohort, the sensitivity of DLRN was 0.981,which also outperformed other three models (0.755[Clinical], 0.925[Radiomics] and 0.943[DTL]). The specificity of DLRN was 0.922 which also high enough to the non-pCR patients. In the test cohort, the sensitivity and specificity of DLRN fell slightly, but still reached 0.812, 0.879,respectively.

Similarly, in LNM, the DLRN achieved much higher AUCs of 0.985 (95% CI,0.970–0.999) and 0.870 (95% CI, 0.774–0.966) in the training and test cohorts. In the training cohort, the sensitivity of DLRN was 0.958, which was the same as Radiomics model, but outperformed other two models (0.681[Clinical], 0.917[DTL]). The NPV of DLRN was 0.975, which mean that the non-LNM patients selected by our model have a high probability of being true negatives.In the test cohort, the sensitivity and NPV of DLRN also higher than other three models.

Table 2

a The performance of models in tumor
Data set	models	AUC	95% CI	Accuracy	Sensitivity	Specificity	PPV	NPV
Train cohort	Clinical	0.834	0.771–0.897	0.784	0.755	0.812	0.580	0.896
	Radiomics	0.979	0.967–0.995	0.933	0.925	0.937	0.845	0.971
	DTL	0.972	0.953–0.991	0.923	0.943	0.915	0.806	0.977
	DLRN	0.984	0.972–0.997	0.938	0.981	0.922	0.825	0.992
Test cohort	Clinical	0.810	0.668–0.952	0.816	0.625	0.909	0.769	0.833
	Radiomics	0.854	0.723–0.986	0.878	0.875	0.879	0.778	0.935
	DTL	0.846	0.695–0.996	0.837	0.875	0.818	0.700	0.931
	DLRN	0.892	0.789–0.996	0.857	0.812	0.879	0.765	0.906

Table 2

b The performance of models in LNM
Data set	models	AUC	95% CI	Accuracy	Sensitivity	Specificity	PPV	NPV
Train cohort	Clinical	0.693	0.625–0.761	0.696	0.681	1.000	0.576	0.789
	Radiomics	0.977	0.958–0.996	0.938	0.958	0.927	0.885	0.974
	DTL	0.953	0.926–0.979	0.882	0.917	0.862	0.795	0.946
	DLRN	0.985	0.970–0.999	0.959	0.958	0.959	0.932	0.975
Test cohort	Clinical	0.738	0.613–0.864	0.735	0.773	1.000	0.680	0.792
	Radiomics	0.779	0.639–0.919	0.796	0.909	0.704	0.714	0.905
	DTL	0.794	0.668–0.919	0.755	0.727	0.778	0.727	0.778
	DLRN	0.870	0.774–0.966	0.776	0.955	0.630	0.677	0.944
Data are presented as percentages except AUC;

To determine the clinical usefulness, decision curve analysis (DCA) was conducted to evaluate the clinical net benefit. Compared with other three models, DLRN had a better clinical benefit in the training and test cohort no matter in terms of tumor or LNM (Fig. 4).

Survival Analysis

To validate the effectiveness of the fusion algorithm and facilitate clinical use, a nomogram was depicted and the statistically significant factors were selected by univariable and multivariate Cox regression model. Our study assessed the utility of the nomogram using C-index. The consistency index (C-index) of the nomogram of the training cohort was 0.761 and the test cohort was 0.731, which showed the good predictive ability of this nomogram. According to the predicted results of the nomogram, patients are divided into high-risk and low-risk groups. The Kaplan-Meier survival curves showed that patients in the high-risk group had significantly poorer overall survival than patients in the low-risk group (P<0.05)(Fig. 5).

In this study, we presented and validated a fusion DLRN model based on longitudinal ultrasound images to personalize prediction the risk of breast tumor residual and LNM invasion after performing NAC for BC patients. The DLRN achieved favorable results in terms of AUC, sensitivity, specificity, and decision curves. Also, the results of survival analysis demonstrated the applicability of the fusion model in clinical practice. Our study confirmed that the fusion model could assist clinicians adjusting treatment regimens for patients, thereby increasing the disease-free survival and overall survival of patient.

The dynamic detection of tumor size, echo, and morphology is routinely used to assess the NAC response, but they are considered relatively late predictors in the NAC process^{[18, 19]}. With the development of ultrasound imaging technology, some studies have reported that traditional ultrasound combined with shear wave elastography or dynamic contrast enhanced ultrasound (CEUS) has improved the predictive ability of NAC treatment response through quantitative parameters. But the diagnostic performance is moderate^{[20, 21]}. Radiomics and deep learning are emerging interdisciplinary combining medical imaging and the computer field, which extracts lots of quantitative information from medical images and shows great potential to assist in clinical diagnosis and treatment^{[11, 22, 23]}. Radiomics and deep learning have been widely used in non-invasive prediction of neoadjuvant chemoradiotherapy responses in various malignant tumors,especially in breast cancer^[24]. Previous studies have evaluated the treatment response using radiomic or deep learning methods based on ultrasound images. Recently, Zhang et al.^[25] constructed a nomogram for predicting the response of patients with axillary lymph node (ALN) positive breast cancer to neoadjuvant chemotherapy (NAC), with AUC value of 0.855. Further, Gu et al.^[16] have proposed two novel DLR models(DLR‐PCR and DLR‐LNM) to assessed the the status of tumors and lymph node metastasis (LNM) independently after NAC, with AUC value of 0.948, 0.896, respectively. From these studies, both traditional radiomics and deep learning models yielded some acceptable results and showed certain advantages in predicting the treatment response after NAC. However, there are few fusion models that package deep learning features with radiomics features and clinical characteristics for the prediction of the status of tumors and LNM in BC patients.

Previous studies have reported that the clinical assessment method based on the images and clinical characteristics has limited performance and clinical practicality^[26]. The reason may be due to the subjectivity and empiricism of humans in determining the effectiveness of NAC, as well as their inability to consistently and accurately identify changes in the tumor micro-environment caused by NAC. The clinical features selected in our study are consistent with previous studies, and the performance of clinical models on each task is limited^[27]. Nevertheless, when we integrating clinical, radiomics, and deep transfer learning features together, the DLRN have achieved much better and satisfactory AUCs for predicting tumor and LNM status, confirming the effectiveness of the fusion model for predicting the status of tumor and LNM in BC patients. This result also reflects that deep learning model may be able to extract high-level abstract features related to tumor micro-environment changes caused by NAC treatment.

The purpose of survival analysis is to analyze the factors that lead to events such as patient death or recurrence within a certain period of time after treatment, and it has important clinical application value^[28].When clinicians have a more accurate assessment of the patient's survival risk, they are more likely to choose appropriate treatment methods^[29]. In this study, we developed a survival model that effectively integrates clinical, radiomics and deep learning data to more accurately predict the survival rate of BC patients, in order to verify the effectiveness of the fusion algorithm and promote clinical application. The results of this study showed that the fusion model can better distinguish high and low risk groups, and has a better ability to predict the survival of breast cancer, which proves the availability of the fusion model in predicting the prognosis of breast cancer.

Our research also has some limitations. Firstly, our study has a small sample size and is a single center study, requiring a larger sample size and multi-center studies to validate our results. Secondly, this is a retrospective analysis. Thirdly, we only used the US data to develop models, and more medical images, including pathological whole slide images and MRI images, might improve our models.

In this study, the fusion model based on US images had better performances for predicting the status of tumor and LNM after NAC in BC patients. At the same time, the fusion model also demonstrated good predictive ability in survival prediction. Therefore, the fusion model is helpful for individual prediction of patients with BC, which could assist doctors to perform preoperative assessment and make better clinical decisions. Future research will construct more larger databases of patients with BC, integrate more clinical factors, further improve the predictive model, and carry out multi-center collaborative research.

NAC Neoadjuvant chemotherapy

BC Breast cancer

LNM Lymph node metastasis

DLRN Deep learning radiomics nomogram

DTL deep transfer learning

ROI Region of interest

ROC Receiver operating characteristic curve

AUC Area under the curve

DCA Decision curve analysis

pCR pathological complete response

Grad-CAM Gradient-weighted Class Activation Mapping

HER2 Human epidermal growth factor receptor 2

ER Estrogen receptor;

PR Progesterone receptor

HR Hormone receptor

NPV Negative predictive value

PPV Positive predictive value

US Ultrasound

Acknowledgements

Some of our experiments were carried out on OnekeyAI platform. Thank OnekeyAI and it's developers' help in this scientific research work.

Authors’ contributions

LYX worked with the LQH and HQH to train the predictive system, analyze the data, provide the figures, and co-wrote the manuscript. SSC provided expertise, reviewed the results and edited the manuscript. SJY coordinated the study plan, obtained and organized the training data. LGL and LH collected the data and performed the initial analyses.All authors read and approved the final manuscript.

Funding

The authors state that this work has not received any funding.

Availability of data and materials

The datasets generated and/or analysed during the current study are not publicly available due [REASON WHY DATA ARE NOT PUBLIC] but are available from the corresponding author on reasonable request.

Ethics approval and consent to participate

Ethics Committee of Guangdong Hospital of Chinese Medicine approval was obtained.

Consent for publication

Not applicable

Competing interests

The authors declare that there have no competing interests.

Author details

¹Department of Ultrasound, The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, China

²Department of Health Management, Zhujiang Hospital, Southern Medical University, Guangzhou, China

³Department of Health Management, Nanfang Hospital, Southern Medical University, Guangzhou, China

^#YueXia Liu, QingHua Liu and QuanHui Hu contributed equally to this manuscript

* Corresponding Author: Department of Ultrasound, Guangdong Hospital of Chinese Medicine, 55 Daxuechengneihuanxi Road, Guangzhou, China

Telephone: 13808826416

e-mail: [email protected]

Zhang W, Xu K, Li Z, Wang L, Chen H (2023) Tumor immune microenvironment components and the other markers can predict the efficacy of neoadjuvant chemotherapy for breast cancer. Clin Transl Oncol. DOI:10.1007/s12094-023-03075-y
Sun S, Mutasa S, Liu MZ et al (2022) Deep learning prediction of axillary lymph node status using ultrasound images. Comput Biol Med. DOI:10.1016/j.compbiomed.2022.105250
Cortazar P, Zhang L, Untch M et al (2014) Pathological complete response and long-term clinical benefit in breast cancer: the CTNeoBC pooled analysis. Lancet. DOI:10.1016/S0140-6736(13)62422-8
Yogananda CGB, Shah BR, Nalawade SS et al (2021) MRI-Based Deep-Learning Method for Determining Glioma Promoter Methylation Status. AJNR Am J Neuroradiol. DOI:10.3174/ajnr.A7029
Abel MK, Greenwood H, Kelil T et al (2021) Accuracy of breast MRI in evaluating nodal status after neoadjuvant therapy in invasive lobular carcinoma. NPJ Breast Cancer. DOI:10.1038/s41523-021-00233-9
Hottat NA, Badr DA, Lecomte S, Besse-Hammer T, Jani JC, Cannie MM (2023) Assessment of diffusion-weighted MRI in predicting response to neoadjuvant chemotherapy in breast cancer patients. Sci Rep. DOI:10.1038/s41598-023-27787-x
Zhao X, Bai JW, Guo Q, Ren K, Zhang GJ (2023) Clinical applications of deep learning in breast MRI. Biochim Biophys Acta Rev Cancer. DOI:10.1016/j.bbcan.2023.188864
Ahn JH, Go J, Lee SJ et al (2023) Changes in Automated Mammographic Breast Density Can Predict Pathological Response After Neoadjuvant Chemotherapy in Breast Cancer. Korean J Radiol. DOI:10.3348/kjr.2022.0629
Trepanier C, Huang A, Liu M, Ha R (2023) Emerging uses of artificial intelligence in breast and axillary ultrasound. Clin Imaging. DOI:10.1016/j.clinimag.2023.05.007
Provenzano E, Brown JP, Pinder SE (2013) Pathological controversies in breast cancer: classification of ductal carcinoma in situ, sentinel lymph nodes and low volume metastatic disease and reporting of neoadjuvant chemotherapy specimens. Clin Oncol (R Coll Radiol). DOI:10.1016/j.clon.2012.10.008
Lambin P, Rios-Velazquez E, Leijenaar R et al (2012) Radiomics: extracting more information from medical images using advanced feature analysis. Eur J Cancer. DOI:10.1016/j.ejca.2011.11.036
Yu FH, Wang JX, Ye XH, Deng J, Hang J, Yang B (2019) Ultrasound-based radiomics nomogram: A potential biomarker to predict axillary lymph node metastasis in early-stage invasive breast cancer. Eur J Radiol. DOI:10.1016/j.ejrad.2019.108658
Yang M, Liu H, Dai Q et al (2022) Treatment Response Prediction Using Ultrasound-Based Pre-, Post-Early, and Delta Radiomics in Neoadjuvant Chemotherapy in Breast Cancer. Front Oncol. DOI:10.3389/fonc.2022.748008
Radak M, Lafta HY, Fallahi H (2023) Machine learning and deep learning techniques for breast cancer diagnosis and classification: a comprehensive review of medical imaging studies. J Cancer Res Clin Oncol. DOI:10.1007/s00432-023-04956-z
Zhou Z, Shin J, Zhang L, Gurudu S, Gotway M, Liang J (2017) Fine-tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally. Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. DOI:10.1109/CVPR.2017.506
Gu J, Tong T, Xu D et al (2023) Deep learning radiomics of ultrasonography for comprehensively predicting tumor and axillary lymph node status after neoadjuvant chemotherapy in breast cancer patients: A multicenter study. Cancer. DOI:10.1002/cncr.34540
Qi TH, Hian OH, Kumaran AM et al (2022) Multi-center evaluation of artificial intelligent imaging and clinical models for predicting neoadjuvant chemotherapy response in breast cancer. Breast Cancer Res Treat. DOI:10.1007/s10549-022-06521-7
Yu FH, Miao SM, Li CY et al (2023) Pretreatment ultrasound-based deep learning radiomics model for the early prediction of pathologic response to neoadjuvant chemotherapy in breast cancer. Eur Radiol. DOI:10.1007/s00330-023-09555-7
Savaridas SL, Sim YT, Vinnicombe SJ, Purdie CA, Thompson AM, Evans A (2019) Are baseline ultrasound and mammographic features associated with rates of pathological completes response in patients receiving neoadjuvant chemotherapy for breast cancer? Cancer Imaging. DOI:10.1186/s40644-019-0251-3
Kim Y, Kim SH, Song BJ et al (2018) Early Prediction of Response to Neoadjuvant Chemotherapy Using Dynamic Contrast-Enhanced MRI and Ultrasound in Breast Cancer. Korean J Radiol. DOI:10.3348/kjr.2018.19.4.682
Ma Y, Zhang S, Zang L et al (2016) Combination of shear wave elastography and Ki-67 index as a novel predictive modality for the pathological response to neoadjuvant chemotherapy in patients with invasive breast cancer. Eur J Cancer. DOI:10.1016/j.ejca.2016.09.031
Yip SS, Aerts HJ (2016) Applications and limitations of radiomics. Phys Med Biol. DOI:10.1088/0031-9155/61/13/R150
Zwanenburg A, Vallières M, Abdalah MA et al (2020) The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology. DOI:10.1148/radiol.2020191145
Park H, Lim Y, Ko ES et al (2018) Radiomics Signature on Magnetic Resonance Imaging: Association with Disease-Free Survival in Patients with Invasive Breast Cancer. Clin Cancer Res. DOI:10.1158/1078-0432.CCR-17-3783
Zhang H, Cao W, Liu L (2023) Noninvasive prediction of node-positive breast cancer response to presurgical neoadjuvant chemotherapy therapy based on machine learning of axillary lymph node ultrasound. J Transl Med. DOI:10.1186/s12967-023-04201-8
Haque W, Verma V, Hatch S, Suzanne Klimberg V, Brian Butler E, Teh BS (2018) Response rates and pathologic complete response by breast cancer molecular subtype following neoadjuvant chemotherapy. Breast Cancer Res Treat. DOI:10.1007/s10549-018-4801-3
Jiang M, Li CL, Luo XM et al (2021) Ultrasound-based deep learning radiomics in the assessment of pathological complete response to neoadjuvant chemotherapy in locally advanced breast cancer. Eur J Cancer. DOI:10.1016/j.ejca.2021.01.028
Shedden K, Taylor JM, Enkemann SA et al (2008) Gene expression-based survival prediction in lung adenocarcinoma: a multi-site, blinded validation study. Nat Med. DOI:10.1038/nm.1790
Ehteshami Bejnordi B, Veta M, Johannes van Diest P et al (2017) Diagnostic Assessment of
Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer. JAMA. DOI:10.1001/jama.2017.14585

Table 1 is available in the Supplementary Files section.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Ultrasound-based deep learning radiomics nomogram for comprehensive prediction of tumor, axillary lymph node status and prognosis of breast cancer patients after neoadjuvant chemotherapy

Status:

Version 1

Abstract

Figures

Introduction

Materials and methods

Patients

Chemotherapy regimens and pathologic analysis

Image Acquisition

Image Segmentation

Deep learning analysis

Radiomics analysis

Feature extraction and selection

Development of the fusion model

Model Evaluation

1. pCR/LNM vs. non-pCR/non-LNM:

2. Survival analysis:

Statistical analysis

Results

Clinical characteristics

Interpretation of Radiomics

Interpretation of Deep Transfer Learning(DTL)

Development and performance of model

Discussion

Conclusion

Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1