Follow the Radiologist: Clinically Relevant Multi-View Cues for Breast Cancer Detection from Mammograms

Automated breast cancer detection using deep learning based object detection models have achieved high sensitivity, but often struggles with high false positive rate. While radiologists possess the ability to analyze and identify malignant masses in mammograms using multiple views, it poses a challenge for deep learning based models. Inspired by how object appearance behaves across multiple views in natural images, researchers have proposed several techniques to exploit geometric correspondence between location of a tumor in multiple views and reduce false positives. We question the clinical relevance of such cues. We show that there is inherent ambiguity in geometric correspondence between the two mammography views, because of which accurate geometric alignment is not possible. Instead, we propose to match morphological cues between the two views. Harnessing recent advances for object detection approaches in computer vision, we adapt a state-of-the-art transformer architecture to use proposed morphological cues. We claim that proposed cues are more agreeable with a clinician's approach compared to the geometrical alignment.

Using our approach, we show a significant improvement of 5% in sensitivity at 0.3 False Positives per Image (FPI) on benchmark INBreast dataset. We also report an improvement of 2% and 1% on in-house and benchmark DDSM dataset respectively. Realizing lack of open source code base in this area impeding reproducible research, we promise to make our source code and pretrained models available publicly.