Evaluation of the Diagnostic Efficacy of the AI-Based Software INF-M01 in Detecting Suspicious Areas of Bladder Cancer Using Cystoscopy Images

Abstract

Background/Objectives: We aimed to evaluate the accuracy of the artificial intelligence (AI)-based software INF-M01 in diagnosing suspected bladder tumors using cystoscopy images. Additionally, we aimed to assess the ability of INF-M01 to distinguish and mark suspected bladder cancer using whole cystoscopy images. Methods: A randomized retrospective clinical trial was conducted using a total of 5670 cystoscopic images provided by three institutions, comprising 1890 images each (486 bladder cancer images and 1404 normal images). The images were randomly distributed into five sets (A–E), each containing 1890 photographs. INF-M01 analyzed the images in set A to evaluate sensitivity, specificity, and accuracy. Sets B to E were analyzed by INF-M01 and four urologists, who marked the suspected bladder tumors. The Dice coefficient was used to compare the ability to differentiate bladder tumors. Results: For set A, the sensitivity, specificity, accuracy, and 95% confidence intervals were 0.973 (0.955–0.984), 0.921 (0.906–0.934), and 0.934 (0.922–0.945), respectively. The mean value of the Dice coefficient of AI was 0.889 (0.873–0.927), while that of clinicians was 0.941 (0.903–0.963), indicating that AI showed a reliable ability to distinguish bladder tumors from normal bladder tissue. AI demonstrated a sensitivity similar to that of urologists (0.971 (0.971–0.983) vs. 0.921 (0.777–0.995)), but a lower specificity (0.920 (0.882–0.962) vs. 0.991 (0.984–0.996)) compared to the urologists. Conclusions: INF-M01 demonstrated satisfactory accuracy in the diagnosis of bladder tumors. Additionally, it displayed an ability to distinguish and mark tumor regions from normal bladder tissue, similar to that of urologists. These results suggest that AI has promising diagnostic capabilities and clinical utility for urologists.