Multi-scale Features Fusion Based Rectal Polyps Detection

The identification of rectal polyps plays a pivotal role in the early detection of rectal cancer. A prevailing challenge in current models employing computer-aided rectal polyp detection lies in the diminished accuracy pertaining to the detection of small objects. To mitigate this challenge, this paper introduces a novel multi-scale detection fusion network model denoted as MF-SSD. This model represents an enhancement over the conventional SSD object detection network. Specifically, our approach incorporates convolutional kernels of varying scales during the feature extraction phase of the SSD network to effectively capture rectal CT features across different scales. This ensures the comprehensive extraction of feature information at diverse target scales. Subsequently, through the fusion of these features across multiple scales, the detection outcomes for rectal polyps are derived. We conducted a series of experiments using publicly available datasets, and the results demonstrated that the proposed MF-SSD model achieved a detection accuracy of 41.2%, surpassing the performance of both R-CNN and SSD models. This outcome underscores the significant performance advantages of the MF-SSD model in rectal polyp detection.