Assisting a complete diagnostic task
Capsule endoscopy produces long recordings in which clinicians must locate relevant intestinal segments, recognise lesions and reach a patient-level diagnosis. INTELCAPE links these steps in a multi-task deep learning system for Crohn’s disease.
The retrospective study used 872 videos from two Chinese hospitals. The system achieved diagnostic AUCs of 0.982 and 0.984 in the two cohorts.
Evaluating clinicians with AI assistance
The study also compared clinicians’ performance with and without assistance. Diagnostic accuracy improved from 76.7% to 94.8%, while mean interpretation time fell from 67.9 to 22.5 minutes.
These task-level measures connect algorithm performance to the work of reading a study. They provide evidence for diagnostic assistance and a basis for subsequent prospective evaluation in clinical practice.
INTELCAPE extends the lab’s clinical AI work beyond colorectal cancer to inflammatory bowel disease. Feng Gao is a co-corresponding author of the study.