Mask-Guided Convolutional Neural Network for Breast Tumor Prognostic Outcome Prediction on 3D DCE-MR Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Breast Cancer Diagnostics

Background:

• Accurate prediction of patient outcomes in breast cancer is crucial for treatment planning.
• Dynamic contrast-enhanced MRI (DCE-MRI) provides valuable information for breast cancer assessment.
• Current deep learning models may not optimally focus on relevant tumor regions in DCE-MRI.

Purpose of the Study:

• To develop and evaluate a mask-guided 3D-CNN architecture for breast cancer patient outcome classification.
• To compare the performance of the mask-guided model against unguided and masked-voxel-only approaches.
• To assess the impact of tumor segmentation masks on the interpretability and accuracy of 3D-CNN predictions.

Main Methods:

• Development of a novel 3D-CNN architecture incorporating tumor masks derived from DCE-MRI.
• Generation and radiologist validation of tumor masks using pre- and post-contrast MRI sequences.
• Classification of patient outcomes, including cancer recurrence and HER2 status, using the developed model.

Main Results:

• The mask-guided 3D-CNN demonstrated superior classification accuracy compared to models using full images or only masked voxels.
• Activation map analysis indicated that the mask guides the model's attention to relevant regions of interest.
• Conservative tumor masks improved model performance over unguided 3D-CNNs, highlighting the benefit of focused attention.

Conclusions:

• Tumor mask-guided 3D-CNNs offer an improved approach for classifying breast cancer patient outcomes from DCE-MRI.
• Incorporating segmentation masks enhances the reliability and interpretability of deep learning models in medical imaging.
• This method provides a more focused and potentially more accurate alternative to unguided deep learning in oncology.