Résumé: Concurrent nuclear segmentation and classification in Hematoxylin & Eosin-stained histopathology images are a crucial task in disease diagnosis and prognosis. Albeit recent advancement of deep learning models, this task remains challenging as each nucleus occupies a limited number of pixels, and nuclei have large intra-class variability and high inter-class similarities in morphology. In this work, we proposed a tissue region-guided dilated Hover-net (TRG-Dilated Hover-net) that consists of a tissue region segmentation model and a dilated Hover-net model. The latter incorporated the dilated convolution and the atrous spatial pyramid pooling feature pyramids to expand the receptive field; therefore, more information about nuclei and their spacial locations can be captured. Our method achieved the state-of-the-art performance on four benchmark datasets of various cancer types and the in-house curated Breast Cancer dataset.

Langue: Anglais