Sosegformer: A Cross-Scale Feature Correlated Network For Small Medical Object Segmentation

Abstract

A mild syndrome with a small infected region is an ominous warning and is foremost in the early diagnosis of diseases. Recently, deep learning algorithms, such as convolutional neural networks (CNN), have been successfully applied to segment natural or medical objects, yielding promising results. However, the analysis of medical objects with small area occupation in images remains largely underexplored. This task poses a significant challenge due to information loss caused by convolution and pooling operations in CNN, particularly for small medical objects. To tackle these challenges, we propose a novel small-object segmentation with transformer (SoSegFormer) network for accurate small-object segmentation in medical images. Quantitative experimental results demonstrate the top-level performance of SoSegFormer, achieving the best mIoU, mDice, MAE, and F2 Score. Notably, it achieved 87.02%, 80.91%, and 65.17% in mDice for segmenting liver tumour, polyp, and sperm objects, which occupy less than 1% of the image areas in ATLAS, PolypGen, and SemSperm datasets.