4T-Net: Multitask deep learning for nuclear analysis from pathology images

Nuclear identification provides nuclear morphology features, distribution, quantity, and other essential features for diagnosing and treating diseases, especially cancer. However, manual nuclear identification is time-consuming with poor sensitivity and low reproducibility. To overcome these limitations, we proposed an automated nuclear identification deep learning-based method using hematoxylin and eosin pathology images. The proposed multitask deep learning model named 4T-Net consists of one encoder and four decoders across different tasks. The proposed method can simultaneously address nuclear segmentation and nuclear classification problems in an end-to-end manner. We evaluated this method on three public datasets: Colorectal Nuclear Segmentation and Phenotypes (CoNSeP), Multiorgan Nuclear Segmentation and Classification (MoNuSAC), and Gastric Lymphocyte Segmentation and Classification (GLySAC). The experimental results confirm the effectiveness and robustness of the four-task network (4T-Net) method in comparison to other competing methods. The contributions of this paper could improve nuclear analysis in computational pathology.