| Resumo: | Of all cancer diseases, breast cancer is the most lethal among women. It has been shown that breast cancer screening programs can decrease mortality, since early detection increases the chances of survival. Usually, a pair of radiologists interpret the screening mammograms, however the process is long and exhausting. This has encouraged the development of computer aided diagnosis (CADx) systems to replace the second radiologist, making a better use of human-experts' time. But CADx systems are associated with high false positive rates, since most of them only use one view (craniocaudal or mediolateral oblique) of the screening mammogram. Radiologist, on the other hand, use both views; frequently reasoning about the diagnosis by noticeable differences between the two views. When considering both projections of a mammogram, lesion matching is a necessary step to perform diagnosis. However this is a complex task, since there might be various lesion candidates on both projections to match. In this work, a matching system is proposed. The system is a cascade of three blocks: candidates detector, feature extraction and lesion matching. The first is a replication of Ribli et al.'s Faster R-CNN and its purpose is to find possible lesion candidates. The second is the feature vector extraction of each candidate, either by using the candidates detector's backbone, handcrafted features or a siamese network model trained for distinguish lesions. The third is the calculus of the distance between feature vector, also using some heuristics to restrain possible non-lesion pairs, and the ranking of the distances to match the lesions. This work provides several options of possible feature extractors and heuristics to be incorporated into a CADx system based on object detectors. The fact that the triplet loss trained models obtained competitive results with the other features extractors is valuable, since it offers some independence between the detection and matching tasks. "Hard" heuristics and "soft" heurisitcs are introduced as methods to restrain matching. The system is able to detect matches satisfactorily, since its accuracy (70%85%) is significantly higher than chance level (30%40%). "Hard" heuristics proposals achieved encouraging results on precision@k, due to its match and candidates exclusion methods, which rejects a significant number of false positives generated by the object detector.
|
|---|