Introduction: Adequate tumor detection is critical in cystoscopy for bladder cancer (BCa) diagnosis, risk stratification and treatment. The adoption of adjunct imaging technologies remains modest. Previously, we developed a deep learning algorithm, CystoNet, that holds the potential to improve performance of white-light cystoscopy (WLC) in a non-invasive and cost-effective manner. Herein we report CystoNet-T which incorporates a transformer-augmented deep learning algorithm to further improve detection of BCa on WLC. Methods: CystoNet-T was developed with a transformer-augmented pyramidal convolutional neural network architecture to improve automated BCa detection during WLC. A training set consisting of 510 tumor-containing frames (54 patients) and a test set of 101 tumor-containing frames (13 patients) were used. Training and test datasets were annotated by a urologist and pathologically confirmed to be BCa. Recall and precision were determined using intersection-over-union as the metric of interest. Harmonic precision-recall mean (F1 score) and average precision (AP) were determined. Performance was then compared against other benchmarks algorithms using the same training and test sets. Results: CystoNet-T detected BCa with a recall of 97.3%, precision of 95.6%, F1 of 96.4% and AP of 91.4% (Figure 1), outperforming the original CystoNet and benchmark models of Faster R-CNN and YOLO by 7.3% in F1 and 3.8% points in AP. Conclusions: We have developed a deep-learning algorithm that improves on prior models in accurately detecting bladder tumors on WLC. Transformer-augmented AI platforms may aid in diagnostic decision making by improving diagnostic yield and target identification. SOURCE OF Funding: NIH R01 CA260426, Urology Care Foundation Research Scholar Award
