University of Twente Student Theses


A 3D deep learning method for the prediction of breast tumor response to neoadjuvant chemotherapy using MR images without the need for a tumor segmentation

Lobeek, Jaap (2022) A 3D deep learning method for the prediction of breast tumor response to neoadjuvant chemotherapy using MR images without the need for a tumor segmentation.

Full text not available from this repository.

Full Text Status:Access to this publication is restricted
Embargo date:9 February 2025
Abstract:Introduction: Pathological complete response (pCR) is confirmed by the absence of residual tumor cells after pathological evaluation of the resected breast specimen after surgery. As about 46% of patients report any form of pain after surgery, and around 15% of the patients are still reporting moderate to severe pain 12 months after surgery, it would greatly improve the quality of life of these patients if pCR can be predicted and futile surgery can be left out in the treatment for these patients. For these reasons, it is desirable to predict pCR after neoadjuvant chemotherapy (NAC) in a non-invasive way. Deep learning approaches have shown promising results to predict the response of breast tumors in breast cancer patients to NAC. However, these approaches need a manual tumor segmentation in the MR images. The segmentation is burdensome and takes much time since the radiologist needs to do this slice by slice. Especially after NAC, segmentation of the tumor is challenging due to unclear tumor boundaries. In this study, it has been investigated to what extent it is possible to use a 3D CNN for predicting pCR without the need for tumor segmentation. For this, three sub-questions have been answered. First, the effect of the size of the region of interest (ROI) on the performance of the 3D CNN has been investigated. Second, it has been investigated which areas in the ROIs are most important for the prediction of the CNN. Lastly, it has been investigated if the usage of the molecular subtype can improve the performance of the CNN. Methods: A large and a small ROI is drawn in the first post-contrast phase of the dynamic contrast-enhanced T1-weighted images, resulting in two datasets. Each ROI dataset is used to train a single-input 3D CNN on post-NAC MR images and a double-input 3D CNN on preand post-NACMR images. Gradient-weighted Class ActivationMapping (Grad-CAM) has been used to visualize the most important regions in the MR images for predicting pCR. An additional input channel for the molecular subtype has been integrated into the single-and doubleinput CNN. Bayesian optimization has been used to find the optimal hyperparameters for all models above. The models are compared in mean accuracy, AUC, and mean Matthews correlation coefficient (MCC) after 10-fold cross-validation. Results: The dataset contained 124 pCR patients and 55 non-pCR patients. The single-input CNN trained on the small ROI achieved the highest performance (a mean AUC, accuracy, and MCC of 0.74§0.09, 69.90§7.77, and 0.22§0.25, respectively). No significant differences in the performance scores were found between training the CNNs with the large or small ROI dataset. Grad-CAM shows that the heart (visible in the large ROIs) and the nipple (visible in both ROIs), influence the prediction of the CNNs. Integration of themolecular subtype did not improve the performance of the CNNs. Discussion and conclusion: It is possible to predict pCR with a 3D CNN without tumor segmentation with an AUC of 0.74 and mean MCC of 0.22. The main complication for all models was localizing the (former) tumor location in the MR images. Parts of the heart and the nipple influenced the performance of the prediction. To improve on the performance of the model, the main recommendations are to use smaller ROIs and to research which MR sequences to use as input for the model.
Item Type:Essay (Master)
NKI-AvL, Amsterdam, The Netherlands
Faculty:TNW: Science and Technology
Subject:44 medicine, 50 technical science in general, 54 computer science
Programme:Biomedical Engineering MSc (66226)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page