Purpose: Development of intelligence systems for bladder lesion detection is cost intensive. An efficient strategy to develop such intelligence solutions is needed.
Materials and methods: We used four deep learning models (ConvNeXt, PlexusNet, MobileNet, and SwinTransformer) covering a variety of model complexity and efficacy. We trained these models on a previously published educational cystoscopy atlas (n = 312 images) to estimate the ratio between normal and cancer scores and externally validated on cystoscopy videos from 68 cases, with region of interest (ROI) pathologically confirmed to be benign and cancerous bladder lesions (ie, ROI). The performance measurement included specificity and sensitivity at frame level, frame sequence (block) level, and ROI level for each case.
Results: Specificity was comparable between four models at frame (range, 30.0%-44.8%) and block levels (56%-67%). Although sensitivity at the frame level (range, 81.4%-88.1%) differed between the models, sensitivity at the block level (100%) and ROI level (100%) was comparable between these models. MobileNet and PlexusNet were computationally more efficient for real-time ROI detection than ConvNeXt and SwinTransformer.
Conclusion: Educational cystoscopy atlas and efficient models facilitate the development of real-time intelligence system for bladder lesion detection.