Semi-Supervised Gastrointestinal Stromal Tumor Detection via Self-Training

The clinical diagnosis of gastrointestinal stromal tumors (GISTs) requires time-consuming tumor localization by physicians, while automated detection of GIST can help physicians develop timely treatment plans. Existing GIST detection methods based on fully supervised deep learning require a large amount of labeled data for the model training, but the acquisition of labeled data is often time-consuming and labor-intensive, hindering the optimization of the model. However, the semi-supervised learning method can perform better than the fully supervised learning method with only a small amount of labeled data because of the full use of unlabeled data, which effectively compensates for the lack of labeled data. Therefore, we propose a semi-supervised gastrointestinal stromal tumor (GIST) detection method based on self-training using the new selection criterion to guarantee the quality of pseudo-labels and adding the pseudo-labeled data to the training set together with the labeled data after linear mixing. In addition, we introduce the improved Faster RCNN with the multiscale module and the feature enhancement module (FEM) for semi-supervised GIST detection. The multiscale module and the FEM can better fit the characteristics of GIST and obtain better detection results. The experiment results showed that our approach achieved the best performance on our GIST image dataset with the joint optimization of the self-training framework, the multiscale module, and the FEM.