256P The synergy of multimodal AI and expert interpretation for improved diagnosis of urinary tract tumors

Abstract

Background: Bladder cancer and upper tract urothelial carcinoma are common urinary tract tumors, whose current diagnostic standards endoscopy and CT are invasive and resource-demanding.As an alternative, cytology and FISH can be used as noninvasive diagnostic tests, but they are less accurate and require manual expert labor.Existing diagnostic AI based on cytology or FISH are unimodal and require cell-level labels that may conflict between cytology and FISH.To overcome this, we developed a semi-supervised, multimodal AI model integrating both.Methods: This retrospective single-center study included 329 patients at Peking University First Hospital (2014-2020) who presented with hematuria or were under post-surgery surveillance.Each patient contributed one cytology and two FISH images.Cytology was classified by Papanicolaou staining (III-V positive, I-II negative).FISH used CSP3, CSP7, CSP17, and GLP P16 probes to detect chromosomal abnormalities (3 abnormal probes defined positivity).Surgical pathology or cystoscopy served as gold standard.AI models (ResNet101, Vision Transformer, CONCH, Dino-Bloom) acted as feature extractors, coupled with a binary classifier.Inputs included cytology, FISH, their combination, and combinations with expert FISH readings.Models were trained with 5-fold cross-validation.Metrics included F1, accuracy, and AUROC, with thresholds set at 0.6. Results:Our results show that cytology or FISH alone yielded moderate performance (AUC 0.61-0.69,F1 0.81-0.85,accuracy 0.69-0.74),though better than manual reading (F1 0.62-0.79,accuracy 0.55-0.69).Combining modalities improved performance (AUC up to 0.74).Cytology+expert FISH readings consistently provided the best outcomes (AUC 0.74, F1 0.85, accuracy 0.74).This suggests that AI is limited by the lack of whole-slide FISH data, whereas human interpretation adds context. Conclusions:In conclusion, multimodal AI integrating cytology and FISH enhances urinary tract tumor diagnosis, and coupling AI with expert FISH reading provides the strongest predictive value, supporting a hybrid AI-clinician workflow for improved accuracy.

Ähnliche Arbeiten