Artificial Intelligence-Powered Cystoscopy Diagnostic Support System: Clinical Application of Multiarchitecture Deep Learning Models

PURPOSE: Cystoscopy is a crucial diagnostic tool in urology for the detection and evaluation of bladder lesions. However, its diagnostic accuracy (ACC) is highly dependent on the operator's experience, which can lead to significant interobserver variability. To address these challenges, this study develops and evaluates two deep learning-based systems for automated cystoscopic image analysis. These systems aim to enhance diagnostic efficiency in clinical settings. MATERIALS AND METHODS: We collected 9362 bladder tumor, cystitis, and postoperative scar images from 2056 examinations for the dataset. We used this dataset to build two deep learning models, one for classifying bladder diseases and the other for locating bladder tumor lesions. We used multiple deep learning model architectures, including convolutional neural networks, Vision Transformer models, and Vision Mamba, to build the models and selected the one with the highest performance. RESULTS: The two deep learning models constructed in this study both demonstrate relatively good capabilities. The common bladder lesion classification model demonstrated high performance, with an area under the curve of 0.872, whereas the bladder tumor segmentation model achieved a Dice similarity coefficient of 0.821 with accurate tumor margin delineation. Both models showed strong generalizability, with the classification model covering most initial diagnostic scenarios. CONCLUSIONS: Our findings indicate that artificial intelligence (AI)-powered cystoscopic analysis significantly enhances both diagnostic ACC and efficiency, aiding clinicians in identifying subtle or complex lesions. This study demonstrates the practicality of deep learning models in real-world clinical applications, laying the groundwork for future AI-driven innovations in urologic diagnostics.