Effect of selection bias on Automatic Colonoscopy Polyp Detection

Abstract

Despite the successful demonstration of the deep learning architectures for Automatic Colonoscopy Polyp Detection (ACPD), studies have argued upon human bias, and lack of interpretability in the existing architectures which hampers its routine clinical and real-time use. To investigate the presence of selection bias, the present work develops a proof-of-concept model with the help of reverse psychology. The model comprises a fine-tuned, interpretable ACPD architecture (VGG-16) and its real-time application (app) called the polyp testing app (link) without consideration of selection bias in the existing open-source colonoscopy datasets. To check the robustness of the developed architecture, it has been trained, validated, and tested on six independent datasets. Evaluation methods in the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) criteria have been used to check the effectiveness of the architecture. Specificity and precision up to 97.3% and 92.0% respectively have been achieved on the validation dataset. A new test set called ‘Gastrolab-Polyp’ has been developed for fair external testing of the developed architecture (link). In comparison to the reference app, the developed app achieved a percentage improvement of 15%, 30.5%, and 14.5% in an overall test dataset accuracy, precision, and recall respectively for the developed Gastrolab-Polyp test dataset. Interpretability methods such as class activation maps, Shapley plots, LIME, and detection analysis of the developed and reference app indicate a bias towards polyp frames. The present work concludes that there is a selection bias of non-polyp frames in five open-source colonoscopy datasets used for polyp detection. Careful dataset collection is needed for its use.