Semi-supervised Gastrointestinal Lesion Segmentation using Adversarial Learning

Accurate segmentation of lesions from gastrointestinal (GI) images is vital for diagnosis and treatment management. However, with the large variability of GI images, it is challenging to enhance GI image lesion segmentation accuracy with small and many unlabeled data. Therefore, this study proposes a semi-supervised lesion segmentation using adversarial learning, which utilized limited labeled and large unlabeled images to train the model. The proposed model is evaluated on GI images from four different sources and by five segmentation evaluation metrics. Finally, the proposed model performs the mean values of segmentation accuracy (DSC=88.56±4.02, IoU=79.69±6.23, Pre=91.55±4.95, Rec=86.07±5.89, and HD=24.85±15.71), which is better than that of existing related works. Thus, a good computer-aided diagnosis system is constructed based on the proposed model for improving accuracy and minimize clinical burdens.