AUTOMATIC DESCRIPTION OF CERVICAL LESIONS IN IVA IMAGES USING ARTIFICIAL INTELLIGENCE

Abstract

This invention relates to an artificial intelligence-based system and method for automated cervical medical image analysis, specifically visual inspection with acetic acid (VIA) images, aimed at supporting early cervical cancer screening. The invention proposes an integrated pipeline consisting of two main stages: detection and segmentation of critical anatomical areas and automatic generation of textual descriptions. In the first stage, the system uses a YOLO-based segmentation model to simultaneously detect and segment cervical areas, columnar epithelium, and lesions. In the second stage, the system applies a Dense-Residual Captioning method combined with Long Short-Term Memory (DRCap-LSTM) to generate textual descriptions for multiple regions within a single image. This approach utilizes residual connections to retain critical visual information and region-based processing to enhance spatial and semantic representation.The resulting system integrates visual information and textual descriptions, comprehensively explaining the characteristics of a clinical area. Thus, this invention not only performs detection and segmentation but also provides text-based interpretation that supports more accurate and efficient clinical decision-making.