Interference-resisted cross-dimensional information aggregation local refinement network for colposcopic image lesion segmentation

In recent years, contour instance segmentation methods have rapidly developed due to advantages such as low computational costs, independence from detection boxes, and the absence of the need to process the internal pixel points of target objects. However, current contour instance segmentation methods still face issues like human-designed rough initial contours and overlapping contour vertex regression paths, which greatly limit the performance of contour instance segmentation algorithm models and increase the difficulty of learning for these models. Additionally, due to objective conditions such as high-light reflection areas and complex lesion edges in colposcopic images, existing models generally exhibit under-segmentation or over-segmentation phenomena. This paper proposes an end-to-end contour instance segmentation algorithm for colposcopic images, based on cross-dimensional attention mechanisms, autonomous initial contour generation, and multi-directional collaborative contour adjustment, effectively reducing the impact of the aforementioned issues and enhancing segmentation accuracy. On the cervical acetic acid map dataset provided by Dr. Yang Xiaohui from the Third Affiliated Hospital of Sun Yat-sen University, this paper's method has achieved improvements of 0.075, 0.047, and 0.013 in the Dice index, IoU index, and AP index respectively compared to the E2ECmodel. On the cervical iodine map dataset, the improvements in each index are 0.089, 0.079, and 0.069 respectively.