Evaluating the Generalizability of CRPU-Net for Polyp Segmentation: Performance on the Piccolo and Real-Time Dataset

Colorectal cancer (CRC) is a global health burden, and its early detection through colonoscopy plays a crucial role in improving patient outcomes. Accurate identification of polyps during colonoscopy is essential for effective diagnosis and treatment. This study replicates the very recently introduced CRPU-Net, a novel deep learning architecture designed in particular for segmenting the polyp present in the colon images automatically. CRPU-Net leverages the strengths of U-Nets allowing the model to extract deeper features and focus on critical regions within the image. We evaluated CRPU-Net's performance on the colonoscopy dataset (CVC-ColonDB) using pixel and object-based metrics. The model achieved a mean accuracy of 96.42%, demonstrating its strong ability to correctly classify pixels. To assess generalizability, the model was tested with various (CVC-ClinicDB, PICCOLO, real-time) datasets. The mean accuracy achieved with the test datasets 96.07 ± 0.17. The low standard deviation between the test datasets and low difference in the accuracy between the validation and test datasets signifying the model's robustness on unseen data. Furthermore, the model achieved a mean F1-Score of 95.44 ± 0.15 and a mean IoU of 93.46 ± 0.42 on test datasets, indicating excellent agreement between predicted and actual polyp boundaries. In conclusion, CRPU-Net demonstrates exceptional performance in colon polyp segmentation, surpassing existing methods. Its high accuracy, sensitivity, specificity, and generalizability suggest its potential as a valuable tool for computer-aided polyp detection in colonoscopy, potentially leading to improved diagnostic accuracy and efficiency.