From WSI-level to patch-level: Structure prior-guided binuclear cell fine-grained detection.

Abstract

Accurate and quick binuclear cell (BC) detection plays a significant role in predicting the risk of leukemia and other malignant tumors. However, manual counting of BCs using microscope images is time consuming and subjective. Moreover, traditional image processing approaches perform poorly due to the limitations in staining quality and the diversity of morphological features in binuclear cell (BC) microscopy whole-slide images (WSIs). To overcome this challenge, we propose a multi-task method inspired by the structure prior of BCs based on deep learning, which cascades to implement BC coarse detection at the WSI level and fine-grained classification at the patch level. The coarse detection network is a multitask detection framework based on circular bounding boxes for cell detection and central key points for nucleus detection. Circle representation reduces the degrees of freedom, mitigates the effect of surrounding impurities compared to usual rectangular boxes and can be rotation invariant in WSIs. Detecting key points in the nucleus can assist in network perception and be used for unsupervised color layer segmentation in later fine-grained classification. The fine classification network consists of a background region suppression module based on color layer mask supervision and a key region selection module based on a transformer due to its global modeling capability. Additionally, an unsupervised and unpaired cytoplasm generator network is first proposed to expand the long-tailed distribution dataset. Finally, experiments are performed on BC multicenter datasets. The proposed BC fine detection method outperforms other benchmarks in almost all evaluation criteria, providing clarification and support for tasks such as cancer screenings.