COLONOSCOPIC POLYP SEGMENTATION USING SEGFORMER-B0 WITH A DICE-BCE HYBRID LOSS

Abstract

Colorectal cancer is one of the leading causes of cancer-related deaths worldwide, with most cases originating from early lesions such as colon polyps. Early detection through colonoscopy is essential to reduce mortality rates; however, accurate polyp identification remains challenging due to variations in shape, size, texture, and illumination conditions. This study aims to implement and evaluate the SegFormer-B0 architecture combined with a Dice-BCE hybrid loss function for polyp segmentation in colonoscopy images. The study utilized the public Kvasir-SEG dataset consisting of 1,000 colonoscopy images with pixel-level annotations. The dataset was divided into 80% training data and 20% validation data. Image preprocessing included resizing to 256×256 pixels and normalization using ImageNet statistics. The model was trained for 25 epochs using the AdamW optimizer with a learning rate of 1×10⁻⁴. Performance evaluation was conducted using Dice Coefficient, Intersection over Union (IoU), Sensitivity, and Specificity metrics. The experimental results demonstrated that the proposed model achieved a Dice Coefficient of 89.92%, Mean IoU of 81.90%, Sensitivity of 89.12%, and Specificity of 98.51%. The training process also showed stable convergence, supported by a training loss of 7.53% and validation loss of 23.30%. The findings indicate that the integration of SegFormer-B0 with the Dice-BCE hybrid loss effectively improves segmentation accuracy and stability while addressing class imbalance issues in colonoscopy images. Therefore, the proposed approach has strong potential to support computer-aided diagnosis systems for colorectal cancer screening.