ConvMixerSeg: Weakly Supervised Semantic Segmentation for CT Liver Images

Abstract

The predictive power of modern deep learning approaches is posed to revolutionize the medical imaging field, however, their usefulness and applicability are severely limited by the lack of well annotated data. Liver segmentation in CT images is an application that could benefit particularly well from less data hungry methods and potentially lead to better liver volume estimation and tumor detection. To this end, we propose a new semantic segmentation model called ConvMixerSeg and experimentally show that it outperforms an FCN with a ResNet-50 backbone when trained to segment livers on a subset of the Liver Tumor Segmentation Benchmark data set (LiTS). We have further developed a novel Class Activation Map (CAM) based method to train semantic segmentation models with image level labels without adding parameters. The proposed CAM method includes a Neighborhood Correlation Enforcement module using Gaussian smoothing that reduces part domination and prediction noise. Additionally, our experiments show that the proposed CAM method outperforms the original CAM method for both classification and segmentation with high statistical significance given the same ConvMixerSeg backbone.