EMTL-Net: Boosting segmentation quality in histopathology images of gland and nuclei by explainable multitask learning network as an optimized strategy

Abstract

In spite of achieving human-level efficacy in gland and nuclei segmentation, modern deep learning-driven techniques still face challenges related to the loss of regional context information and disregard the importance of long-range semantic information during the optimization. Furthermore, one typical issue with CNN-based approaches is also their lack of clinical explainability. In such context, we propose an explainable multitask learning network (EMTL-Net) to enhance the segmentation quality of gland and nuclei segmentation via histopathology images. Initially, to obtain better feature maps, boost the network growth, and reduce the computational cost in the encoding path, we introduce the tweaked residual module (TRM) in which the diagnosis information of histopathology images structure will not be overlooked. After that, we design the two modules to address the challenges of regional context information, such as the complementary feature detection module (CFDM) for the intermediate layers and the cascade feature module (CFM) for the deeper layer of the encoder block. We also design a competitive decoder by casting an efficient pyramid split attention (EPSA) unit that preserves long-range semantic information, stimulates robust features, and dynamically recalibrates spatial and channel-wise features. Extensive experimental findings on real-world datasets, including Glas, MoNuSeg and TNBC, reveal that our proposed model has achieved a Dice Coefficient and Intersection over Union of 0.923 and 0.869 on Glas, 0.855 and 0.741 on TNBC, and 0.829 and 0.713 on MoNuSeg when compared with the SOTA approaches for gland and nuclei segmentation. In addition, the suggested EMTL-Net is also examined for paying full attention on the regions of the gland and nuclei to meet the needs of clinical scenarios.

Keywords

• Gland and nuclei segmentation
• Automated diagnosis
• Explainable multitask learning
• Histopathology images
• Deep learning