SFOD-Trans: semi-supervised fine-grained object detection framework with transformer module

As the labeling cost of object detection for medical images is very high, semi-supervised learning methods for medical images are investigated. In this paper, semi-supervised fine-grained object detection framework with transformer module (SFOD-Trans) is proposed for hepatic portal vein detection. It adopts Sparse R-CNN as the backbone. In detection model, the transformer module is introduced and contrastive loss is added to improve the performance of fine-grained object detection. In order to complete the information transfer both of labeled and unlabeled pictures, a new fusion module named normalized ROI fusion (NRF) is designed based on the characteristics of hepatic portal vein. We run a large number of experiments on a dataset of 1000 real CT scans. The results show that Average Precision (AP) and Average Recall (AR) of the proposed method reach 0.773 and 0.831 respectively with the 300 labeled and 1500 unlabeled samples. An overview of semi-supervised fine-grained object detection framework with transformer module (SFOD-Trans). There are two parallel branches to train supervised loss and semi-supervised loss respectively.