Classes U-Net: A method for nuclei segmentation of photoacoustic histology imaging based on information entropy image classification

Abstract

Due to its ability to reveal the intrinsic molecule specificity of DNA/RNA at the subcellular lateral resolution, photoacoustic (PA) microscopy holds great promise in histopathology imaging of tissue samples. An essential marker for subsequent illness study and diagnosis is the histopathological picture. Segmenting the histopathological image of cell nuclei has been significantly aided by contemporary image processing technology, while they usually suffer from inadequate segmentation or training resource waste. There are many traditional methods, such as threshold segmentation and region-growing segmentation, but these traditional segmentation methods are affected by the gray distribution of the image itself, so the accuracy of segmentation is difficult to meet the requirements of the index. To address this challenge, we propose an approach, called Classes U-Net, which combines the information entropy classification with U-Net and U-Net++ architecture for the segmentation of photoacoustic histology images. The results show that our Classes U-Net effectively improves the DICE to 91.43 %, IOU to 84.215, better than U-Net’s 83.28 % and 71.35 %, better than U-Net++’s 84.60 % and 73.31 %, and our Classes U-Net reduce the required computing resources.