ReCasNet: reducing mismatch within the two-stage mitosis detection framework

Abstract

Mitotic count (MC) is an important histological parameter for cancer diagnosis and grading, but the manual process to obtain this metric is tedious and not fully reproducible across different pathologists. To mitigate this problem, several deep learning models have been utilized to speed up the process. Typically, the problem is formulated as a two-stage deep learning pipeline: the detection stage for proposing the potential candidates for mitotic cells and the classification stage for refining prediction confidences from the former stage. However, this paradigm can lead to inconsistencies in the classification stage due to the poor prediction quality of the detection stage and the mismatches in training data distributions between the two stages. This thesis proposes a Refine Cascade Network (ReCasNet), an improved deep learning pipeline that introduces three improvements to alleviate the aforementioned problems. First, window relocation was used to suppress poor-quality false positive boxes produced by the detection stage around the sliding window border. Second, we proposed an additional deep learning model to align the poorly centered objects to the true object center. Third, additional data were queried from the training slides to teach the classification stage to bridge the training distribution gap between the two stages. We evaluated the performance of ReCasNet on two public large-scale mitotic figure recognition datasets, canine cutaneous mast cell tumor (CCMCT) and canine mammary carcinoma (CMC). By using our proposed pipeline, we achieved up to 4.8% F1 improvements for mitotic cell detection performance and 44.1% reductions in mean absolute percentage error (MAPE) for MCprediction. Techniques that underlie our proposed method can be generalized to other detection and classification algorithms and should contribute to improving the performances of deep learning models in broad digital pathology applications.