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

Finished: 2022-02-09

MSc assignment

Over 50% of women diagnosed with breast cancer require chemotherapy as part of their treatment. This is increasingly administered before (neoadjuvant chemotherapy = NAC) rather than after surgery of the breast. A benefit of NAC is the possibility of breast-conserving surgery due to the downsizing of the tumor load. Depending on the subtype of breast cancer, in 10 to 80% of patients, no residual tumor (pathologic complete response = pCR) is found after NAC. Currently these patients however still receive surgery, since the selection of patients with pCR is unreliable. As about 25% of patients suffer from pain and/or deformation of the breast after breast surgery, it would greatly improve the quality of life if surgery could be omitted in these patients. For these reasons, it is desirable to identify which patients will respond to NAC as early as possible.

Different imaging modalities can be used to predict tumor response to NAC in breast cancer patients, such as mammography, ultrasound, PET/CT, and breast MRI. Breast MRI is considered the most accurate imaging modality for response prediction. However, the accuracy of the radiologist's visual assessment in predicting tumor response to NAC remains insufficient to adapt treatment in clinical practice. Radiomics-based approaches, involving the extraction and analysis of quantitative imaging features to determine relationships between the features and the underlying pathophysiology, have shown promise in the assessment of tumor response in retrospective studies, but with no prospective validation.

More recently, deep learning approaches have emerged as a promising tool for response prediction. Convolutional neural networks can discover visual patterns in images. These visual patterns can be used by the algorithm as imaging features for breast cancer diagnosis, subtype classification, diagnosis for metastasis, or, in this case, for predicting response before and early in NAC. At the NKI, patients with breast cancer are treated with chemotherapy before surgery. Accurate imaging of the breast during and following NAC is crucial for the optimal selection of patients suitable for surgical resection and preparation of a surgical plan. During NAC, lesions may show complete response histologically, which is only possible to validate after surgery by histopathology analysis of the excised specimen. Meaning that these patients had unnecessary surgery.

To improve on the assessment of NAC response prediction, the Image-Guided Surgery group in the NKI aiming to design and develop a machine/deep learning algorithm that can predict tumor response to NAC by analyzing pre-NAC and post-NAC MRI images. To achieve this, the following goals have been set for this project:
• Developing an algorithm for automatic detection of the region of interest in breast MRI images (lesion localization).
• Designing and developing a multi-channel neural network to predict response to chemotherapy using both pre-NAC and post-NAC MRI images.
• Visualizing the attention of such a deep learning network, to localize the residual tumor in breast MRI scans in patients with no complete response.
• Achieving an expected AUC of >0.97 for response prediction, followed by a prospective validation.
• Investigating the importance of pre-NAC and post-NAC MRI scans individually in response prediction.

For this project, the NKI has a collection of around 200 pre and post-NAC MRI images available to train and validate an algorithm. Also, a medical Ph.D. student has just started collecting more MRI data, this additional data can also be used in this project.