A novel artificial intelligence segmentation model for early diagnosis of bladder tumors.

Objective: Despite cystoscopy plays an important role in bladder tumors diagnosis, it often falls short in flat cancerous tissue and minuscule satellite lesions. It can easily lead to a missed diagnosis by the urologist, which can lead to a swift tumor regrowth following transurethral resection of the bladder tumor (TURBT). Therefore, we developed a deep learning-based intelligent diagnosis system for early bladder cancer to improve the identification rate of early bladder tumors.

Methods: Video data from 273 bladder cancer patients who underwent TURBT at Zhongnan Hospital were collected. The dataset was carefully annotated by urologists to clearly define tumor boundaries. Subsequently, we developed a new bladder tumor segmentation network (BTS-Net) based on transformer to accurately diagnose early-stage bladder cancer lesions.

Results: Our experiments demonstrate that the BTS-Net we developed has outperformed other method on the external B validation dataset, achieving a MPrecision of 91.39%, a MRecall of 95.71%, a MIoU of 88.18% and an F1-score of 93.18%. The BTS-Net showed high accuracy with real-time processing speed at 23 fps.

Conclusion: Missed detection of satellite lesions in early bladder tumors often leads to tumor recurrence. Our BTS-Net is capable of segmenting all potential satellite lesions in surgical videos, without the need for complex professional equipment. This AI-assisted diagnosis system has the potential to improve surgical outcomes by ensuring comprehensive treatment of all tumor-related areas during TURBT.