A novel approach to overcome black box of AI for optical diagnosis in colonoscopy.

Accurate real-time optical diagnosis that distinguishes neoplastic from non-neoplastic colorectal lesions during colonoscopy can lower the costs of pathological assessments, prevent unnecessary polypectomies, and help avoid adverse events. Using a multistep process, this study developed an explainable artificial intelligence method, niceAI, for classifying hyperplastic and adenomatous polyps. Radiomics and color were extracted, followed by feature selection with deep learning features using Spearman's correlation analysis. The selected deep features were merged with the narrow-band imaging International Colorectal Endoscopic grading, aligning with endoscopists' decision-making process to produce an interpretable diagnostic output. Initially, 2,048 deep features were identified; these were reduced to 103 in the second screening, and finally to 14. Similarly, 24 radiomics features were selected, whereas no color features were chosen. Comparative evaluation showed that niceAI had accuracy comparable to that of deep learning models (area under the curve, 0.946; accuracy, 0.883; sensitivity, 0.888; specificity, 0.879; positive predictive value, 0.893; negative predictive value, 0.872). This study introduces a novel system that combines radiomics and deep features to enhance the transparency and understanding of optical diagnosis. This approach bridges the gap between artificial intelligence predictions and clinically meaningful assessments, thereby offering a practical solution for enhancing diagnostic accuracy and clinical decision-making.