An Integrated Deep Architecture for Lesion Detection in Breast MRI

Complex nature of medical images and tedious process of data exploration calls for the development of Computer Aided Detection (CADe) methods to ease the process of lesion detection. Recent deep learning-based object detectors from computer vision are adapted to the creation of CADe lesion detectors. This research starts with state-of-the-art object detectors, namely Faster R-CNN, YOLO v2 and Grad-CAM, to determine the location of lesions in Magnetic Resonance Images of breast. A series of experiments are conducted to find the best set up for maximizing the Average Precision (AP) of each method. Consequently, AP values of 0.6993 and 0.7651 are obtained for Faster R-CNN and YOLO v2 respectively. Taking into consideration the pros and cons of each method, we propose different integration architectures in order to overcome the shortcomings of each algorithm, hence enhancing the overall lesion detection performance. The integrated architectures succeed to obtain an AP value up to 0.8097 while providing explainable reasoning that is essential for medical CADe.