Comparative analysis of machine learning frameworks for automatic polyp characterization

Early detection and characterization of adenomatous polyps, the precursors to colorectal cancer is crucial for effective diagnosis and treatment of colorectal cancer. A suitably trained machine learning algorithm can perform automatic real-time optical characterization of polyps during colonoscopy procedures with virtual chromoendoscopy systems, such as Narrow Band Imaging (NBI). This study proposes to perform a detailed comparative analysis between the performances of two popular machine learning frameworks, namely training from scratch and transfer learning, using three state-of-the-art architectures ResNet50, EfficientNetB2, and ViTb_32 with one indigenous dataset and one publicly available dataset. The indigenous dataset was created using 1739 NBI frames extracted from colonoscopy videos of 30 polyp masses from 24 patients [870 hyperplastic polyp frames and 869 adenomatous polyp frames]. The publicly available dataset also consisted of NBI frames. Blurry frames or frames with specular reflections were not excluded to closely resemble reallife polyp characterization in a clinical setting. Among the three models used in two different frameworks, EfficientNetB2 in the training from scratch framework yielded the best accuracy, AUC and NPV of 91.91%, 0.99 and 89.06%. The results show that the training from scratch framework outperformed the transfer learning framework in terms of the evaluation metrics. To the best of our knowledge, this is the first indigenous study to focus on polyp classification using solely Narrow Band Imaging (NBI) polyp images. Our results using only NBI polyp frames outperformed those obtained with a combination of white light and NBI frames, as observed in similar existing studies.