According to current guidelines, the diagnosis of lymph node metastasis in bladder cancer is performed through pathological assessment, which, however, is time consuming and not devoid of limitations.
In a study published in The Lancet Oncology, the authors developed an artificial intelligence (AI)-based model to detect lymph node metastases (including micrometastases) from whole slide images in patients with bladder cancer. In this retrospective, multicentre, diagnostic study, 998 patients with bladder cancer who had undergone cystectomy and pelvic lymph node dissection across five hospitals in China were included. A total of 7,991 images from these patients were used to develop the lymph node metastases diagnostic model (LNMDM): images from 831 patients from two hospitals were divided between the training and the internal validation set, whereas images from 167 patients from the other three hospitals were used as external validation sets. In the five validation sets, the area under the curve (AUC) ranged from 0.978 (95% CI 0.960–0.996) to 0.998 (95% CI 0.996–1.000). For the model validation, slides were assessed by six pathologists, three junior pathologists (who had completed a 3-year residency programme) and three senior pathologists (who had >10 years of experience). Pathologists assessed the slides without AI assistance (non-AI mode) or after suspicious areas in the slide had been highlighted by AI (AI-assisted mode). LNMDM showed a higher sensitivity (0.983 (95% CI 0.941–0.998)) and a lower specificity (0.925 (95% CI 0.862–0.965)) than both junior and senior pathologists (sensitivity: 0.906 (95% CI 0.871–0.934) and 0.947 (95% CI 0.919–0.968), respectively; specificity: 0.967 (95% CI 0.942–0.983) and 0.989 (95% CI 0.972–0.997), respectively). The sensitivity of pathologists’ assessment improved with AI assistance for both junior (0.953) and senior (0.986) pathologists. The authors found that LNMDM would enable the removal of 80–92% of negative slides from pathologist’s evaluation without reducing sensitivity. Moreover, the mean review time per slide was substantially reduced compared with non-AI mode (−23.3% and −22.5% for junior and senior pathologists, respectively).
This is a preview of subscription content, access via your institution
