Deep learning-based automated identification of papillary bladder cancer from endoscopic still images using SE-ResNeXt-50.

Bladder cancer remains a major global health challenge, necessitating rigorous endoscopic surveillance and the precise identification of neoplastic lesions. While the detection of overt tumors is routine, achieving high specificity and distinguishing borderline tumor structures are critical for optimal clinical decision-making. This is especially vital given the rapid evolution of intravesical therapies for BCG-unresponsive non-muscle invasive bladder cancer (NMIBC). Recent clinical advancements with novel agents-such as gemcitabine-releasing intravesical systems (TAR-200), nadofaragene firadenovec, and the IL-15 superagonist complex (N-803)-fundamentally require a clinically "tumor-free" state through complete resection to ensure maximum therapeutic efficacy. This study presents an advanced artificial intelligence (AI) system developed, utilizing a SE-ResNeXt-50 architecture to automate the classification of papillary bladder tumors from cystoscopic still images. To ensure transparency, Grad-CAM analysis was employed, revealing that the model prioritizes key morphological tumor features. Although background artifacts like air bubbles or mesh-like structures occasionally influence the signal, the system maintains high diagnostic reliability. Our findings suggest that this deep learning framework provides robust decision support, facilitating the rigorous complete resection necessary for successful intervention with next-generation intravesical agents.