Automatic textual description of colorectal polyp features: explainable artificial intelligence based on the BASIC classification

Aims Computer-aided diagnosis (CADx-)systems could improve optical diagnosis of colorectal polyps (CRPs) by endoscopists. For integration into clinical practice, better understanding of artificial intelligence (AI) is needed. A branch of deep learning and explainable AI is automatically generating textual descriptions from images to improve understanding. We aimed to develop a CADx-system generating automatic textual descriptions for CRPs based on Blue Light Imaging (BLI) Adenoma Serrated International Classification (BASIC)[1].

Methods Training data contained 35 hyperplastic polyps, 12 sessile serrated lesions (SSLs) and 48 adenomas, with 6525 corresponding textual descriptions by endoscopists. Testing data contained 15 hyperplastic polyps, three SSLs, 36 adenomas, and one colorectal carcinoma. Both databases consisted of High Definition White Light (HDWL), BLI, and Linked Color Imaging (LCI) images. CADx’s 165 generated descriptions were compared to 1857 descriptions from nineteen endoscopists. References not matching histological diagnoses were excluded. The Recall Oriented Understudy for Gisting Evaluation Longest common subsequence (ROUGE-L) score measured the longest word segment in generated descriptions corresponding with reference descriptions.

Results A CADx-system generating automatic textual descriptions of CRP features was successfully developed ([Figure 1]). ROUGE-L scores (%) per category were: Complete sentence 83%, BASIC descriptors 70%, Morphology & size 89%, Surface 92%, Pit pattern 85%, and Vessels 59%