ADVANCING COLORECTAL POLYP DETECTION: AN AUTOMATED SEGMENTATION APPROACH WITH COLRECTSEG-UNET

Abstract

Purpose: This study aimed to achieve two primary goals. First, we sought to develop a lightweight convolutional neural network (CNN) model, COLRECTSEG-UNet, for the purpose of automated polyp segmentation and localization. Second, we evaluated the generalizability of the proposed model by testing it on a proprietary colonoscopy image dataset (own dataset) in addition to a publicly available benchmark (CVC-ColonDB and Kvasir dataset). Methods: The COLRECTSEG-UNet architecture was employed to perform polyp segmentation. Subsequently, the model was trained and validated using an 80:20 data split from the CVC-ColonDB dataset. The efficacy of these data split was confirmed through [Formula: see text]-fold cross-validation. To elucidate the significance of each layer within the COLRECTSEG-UNet design, an ablation study was conducted. The performance of the model was assessed using various metrics on the CVC-ClinicDB, own dataset and Kvasir dataset during the testing phase. These metrics were then compared to those achieved by current state-of-the-art models and findings reported in recent literature. Results: When evaluated on the CVC-ClinicDB dataset, the COLRECTSEG-UNet model achieved outstanding accuracy of 0.9591, intersection over union (IoU) of 0.9299 and F1-Score of 0.9530. Similarly, impressive results were obtained on the own dataset, with accuracy, IoU and F1-Score values of 0.955, 0.9346 and 0.9539, respectively. Additionally, Kvasir dataset attained the accuracy, IoU and F1-Score of 0.9542, 0.9291 and 0.9565, respectively. These results demonstrate that the proposed COLRECTSEG-UNet outperforms the existing benchmark models and surpasses the performance reported in current literature. Conclusion: The implementation of COLRECTSEG-UNet as a lightweight model paves the way for its integration as a backend component within medical imaging system software. This has the potential to significantly aid gastroenterologists in clinical decision-making during interventions.