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

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• High labeling costs in medical imaging hinder object detection development.
• Semi-supervised learning offers a promising solution for data-scarce medical image analysis.

Purpose of the Study:

• To propose a novel semi-supervised fine-grained object detection framework (SFOD-Trans) for accurate hepatic portal vein detection.
• To address the challenge of limited labeled data in medical image object detection.

Main Methods:

• The SFOD-Trans framework utilizes Sparse R-CNN as a backbone.
• Incorporates a transformer module and contrastive loss for enhanced fine-grained detection.
• Introduces a normalized ROI fusion (NRF) module for effective information transfer between labeled and unlabeled data.

Main Results:

• Achieved an Average Precision (AP) of 0.773 and Average Recall (AR) of 0.831 on a dataset of 1000 CT scans.
• Demonstrated strong performance with only 300 labeled and 1500 unlabeled samples.
• The transformer module and NRF module significantly boosted detection performance.

Conclusions:

• The proposed SFOD-Trans framework effectively improves semi-supervised fine-grained object detection in medical images.
• This approach offers a cost-effective solution for hepatic portal vein detection using limited labeled data.
• SFOD-Trans shows potential for broader applications in medical image analysis.