Deep Learning and Minimally Invasive Inflammatory Activity Assessment: Development and Score Correlation of a Pan-Endoscopic Convolutional Network

Introduction: Capsule endoscopy (CE) is a valuable tool for assessing inflammation in Crohn's (CD) patients. The current standard for evaluating inflammation are validated scores (and clinical laboratory values) like the Lewis score (LS), Capsule Endoscopy Crohn's Disease Activity Index (CECDAI), and ELIAKIM score. Recent advances in artificial intelligence (AI) have automatically made it possible to select the most relevant frames in capsule endoscopy. In this study, we aimed to develop an automated scoring system using CE images to grade inflammation objectively. Methods: Pan-enteric CE videos (PillCam Crohn's) performed in CD patients between 09/2020 and 01/2023 were retrospectively reviewed, and LS, CECDAI, and ELIAKIM scores were calculated. Our group developed a convolutional neural network-based automated score consisting of the percentage of positive frames selected by the algorithm (for small bowel and colon separately). We correlated clinical data and the validated scores with the artificial intelligence-generated score (AIS). Results: A total of 61 patients were included. The median SL was 225 [0-6,06], CECDAI was 6 [0-33], ELIAKIM was 4 [0-38], and SB_AIS was 0.5659 [0-29.45]. We found a strong correlation between SB_AIS and LS, CECDAI, and ELIAKIM scores (Pearson's r= 0.751, r= 0.707, r= 0.655, p = 0.001). We found a strong correlation between SL and ELIAKIM (r= 0,768, P = 0.001) and a robust correlation between CECDAI and SL scores (r= 0,854, p = 0.001) and CECDAI and ELIAKIM (r= 0,827, P = 0.001) (Figure 1). Conclusion: Our study showed that the AI-generated score strongly correlated with validated scores, indicating that it could serve as an objective and efficient method for evaluating inflammation in CD patients. As a pre-proof study, our findings provide a promising basis for subsequent refining of a CE score that can accurately correlate with prognostic factors and aid in managing and treating CD patients.Figure 1.: A - Output obtained from the application of the convolutional neural network. A blue bar represents a correct prediction. Red bars represent an incorrect prediction. The category with the highest probability was outputted as the CNN’s prediction. B - Evolution of the accuracy of the convolutional neural network during training and validation phases, as the training and validation datasets were repeatedly inputted in the neural network. C - ROC analyses of the network’s performance. Abbreviations: CNN: convolutional neural network; N: normal mucosa; PUE: ulcers and erosions of the enteric and colonic.