Aim: Colorectal cancer is a leading cause of cancer-related mortality, emphasising the need for accurate polyp segmentation during colonoscopy for early detection. Existing methods often struggle to generalize effectively across diverse clinical scenarios. This study introduces DeepPolyp, an artificial intelligence framework designed for comprehensive benchmarking and real-time clinical deployment of polyp segmentation models. Methods: Transformer-based segmentation models, SegFormer and SSFormer, were trained from scratch using an extensive dataset comprising public collections (CVC-ClinicDB, ETIS-LaribPolypDB, Kvasir) and recently augmented datasets (PolypDataset-TCNoEndo, PolypGen). Training involved standardized data augmentation, learning rate schedules, and early stopping. Models were evaluated using Dice and Intersection over Union (IoU) metrics. Real-time inference performance was assessed on an NVIDIA Jetson Orin device with ONNX and TensorRT optimizations. Results: SegFormer-B4 achieved the highest accuracy (Dice: 0.9843, IoU: 0.9694), but was not selected for clinical deployment due to computational constraints. SegFormer-B2 provided comparable accuracy (Dice: 0.9787, IoU: 0.9588) with significantly faster inference (94 ms per frame), offering an optimal balance suitable for real-time clinical use. SSFormer showed lower accuracy and slower inference, limiting its practical deployment. Conclusions: DeepPolyp enables systematic evaluation of polyp segmentation models, assisting in selecting models based on both performance and computational efficiency. Despite superior accuracy from SegFormer-B4, SegFormer-B2 was selected for clinical deployment due to its advantageous balance between accuracy and real-time execution efficiency.
Mameli et al. (Mon,) studied this question.