Computer-aided diagnosis system using only white-light endoscopy for the prediction of invasion depth in colorectal cancer.

BACKGROUND AND AIMS: Endoscopic treatment is recommended for low-grade dysplasia (LGD), high-grade dysplasia (HGD), and colorectal cancer (CRC) with submucosal (SM) invasion <1000 μm. However, diagnosis of invasion depth requires experience and is often difficult. This study developed and evaluated a novel computer-aided diagnosis (CAD) system to determine whether endoscopic treatment is appropriate for colorectal lesions using only white-light endoscopy (WLE).

METHODS: We extracted 3442 images from 1035 consecutive colorectal lesions (105 LGDs, 377 HGDs, 107 CRCs with SM <1000 μm, 146 CRCs with SM ≥1000 μm, and 300 advanced CRCs). All images were WLE, nonmagnified, and nonstained. We developed a novel CAD system using 2751 images; the remaining 691 images were evaluated by the CAD system as a test set. The capability of the CAD system to distinguish endoscopically treatable lesions and untreatable lesions was assessed and compared with the results from 2 trainees and 2 experts.

RESULTS: The CAD system distinguished endoscopically treatable from untreatable lesions with 96.7% sensitivity, 75.0% specificity, and 90.3% accuracy. These values were significantly higher than those from trainees (92.1%, 67.6%, and 84.9%; P < .01, <.01, and <.01, respectively) and were comparable with those from experts (96.5%, 72.5%, and 89.4%, respectively). Trainees assisted by the CAD system demonstrated a diagnostic capability comparable with that of experts.

CONCLUSIONS: The CAD system had good diagnostic capability for making treatment decisions for colorectal lesions. This system may enable a more convenient and accurate diagnosis using only WLE.