Portal area segmentation is an important step in the quantitative histological analysis process for hepatitis grading. However, portal areas often appear of different sizes and appearances due to the variations of surrounding components such as the ductule, bile duct, artery, and portal vein. The slim fibrosis expanding from the portal area further increases challenges of the portal area segmentation. A Multi-scale Receptive Field Convolutional Neural Network (MRF-CNN) for the segmentation of the liver portal areas in hematoxylin and eosin (HE) stained whole slide images (WSIs) is proposed in this paper. The MRF-CNN adopts the atrous spatial pyramid pooling (ASPP) with multiple atrous rates and symmetric encoder-decoder with feature concatenation architecture. The atrous rates in ASPP are devised of receptive fields to extract features of meaningful tissue components in parallel in portal areas. Along with the MRF-CNN, a small object sensitive loss function is also proposed to have the network focus on small portal areas and slim fibrosis. The results show that the proposed model achieves Intersection over Union (IOU) of 0.87 and sensitivity of 0.92. Compared to recent segmentation researches such as Fully Convolutional Network (FCN), U-Net and SegNet, the proposed network achieves an overall the best IOU and sensitivity performance. Experimental results also show that the designed ASPP block benefits in feature extraction, and the ability of identifying small objects in proposed small object sensitive loss has a significant improvement of the segmentation result comparing to the original cross entropy loss.
|頁（從 - 到）||623-634|
|期刊||IEEE Journal on Emerging and Selected Topics in Circuits and Systems|
|出版狀態||Published - 2019 12月|
All Science Journal Classification (ASJC) codes