Domain Adaptive Small Polyp Detection Using Image Enhancement and Feature Fusion

Automated gastric polyp detection has been proven crucial for improving diagnostic accuracy. However, when data is subject to domain shifts (e.g., different hospital detection equipment and imaging protocols), deep learning-based detection methods fail to perform effectively. Unsupervised domain adaptation (UDA) has been demonstrated as an effective solution to this problem. Nevertheless, existing UDA detection methods often struggle with the issue of small polyp information loss in gastric polyp images. In this paper, a Domain Adaptive Small Polyp Detection(DASP) method is designed. This method includes a small-sized information enhancement module and an adaptive feature fusion module to address this issue. The small object information enhancement module employs the Sobel operator and parallel connections of multi-scale convolutional kernels, thereby enhancing both the texture and semantic information of small objects. The adaptive feature fusion module utilizes a multi-receptive field attention module designed by us to adaptively generate fusion weights, integrating the enhanced small object information into the corresponding feature layers. Experimentally, we compared our proposed method with advanced domain adaptation object detection methods such as SIGMA++ and HTCN. The experimental results indicate that our method significantly outperforms other domain adaptation detectors in the task of gastric polyp detection.