Deep learning-assisted detection of lymph node metastases in bladder cancer.

BACKGROUND: Bladder cancer is the most common malignancy of the urinary tract and is often treated with radical cystectomy with lymph node dissection, followed by a thorough pathological evaluation of the dissected nodes. This is crucial for both accurate prognosis and successful treatment. However, histopathological lymph node examination is labor-intensive and time-consuming. AIM: To develop a supervised deep learning model for detecting nodal metastases in bladder cancer, and to evaluate its influence on pathologists' assessment efficiency. METHODS: This study included 100 histopathological slides representing lymph nodes collected between 2015 and 2024 from the Sahlgrenska University Hospital and scanned into whole slide images, divided into training and validation/test sets. The algorithm was trained with 3437 pixelwise annotations. In the validation set, 50 regions containing normal tissue and metastasis were assessed by AI and two specialist pathologists as validators. In the test set, two pathologists reviewed entire slides and measured review times without, and then, after the washout period, with AI assistance. RESULTS: In the test set, the AI model achieved a sensitivity of 100% and a specificity of 60% for the detection of metastases. For pathologist 1, the median review time decreased from 9 s to 4.2 s with AI assistance, (sign test: Z = -4.73, p < 0.001). For pathologist 2, the median review time decreased from 12 to 4 s (sign test: Z = -5.00, p < 0.001). CONCLUSIONS: This study demonstrates that an AI model trained on a limited dataset can effectively assist pathologists in the detection of lymph node metastases in bladder cancer while significantly reducing review time. These findings suggest that institutions with limited case volumes may be able to develop and implement locally trained AI models to support routine histopathological assessment.