In the multicentre IP8-FLUORESCE study, a deep learning model for cancer image analysis flagged malignant fluorescence confocal microscopy (FCM) images during radical prostatectomy with an internal-test AUC of 0.93, 87.5% sensitivity, and 97.9% specificity; on external validation, AUC was 0.83 with 91.3% sensitivity. The model was developed on 275 images from 24 patients and tested externally on 46 independent images, using histopathology as the reference standard.

Why It Matters To Your Practice

• Intraoperative margin assessment in prostatectomy is time-sensitive and often depends on expert pathology support that may not be immediately available.

• An AI-assisted FCM workflow could help identify malignant neoplasm at the surgical margin in real time, potentially supporting faster decisions in the operating room.

• Strong negative predictive performance may be especially useful when clinicians need added confidence that a margin is benign.

Clinical Implications

• Internal testing showed sensitivity of 87.5% and specificity of 97.9%; external validation showed sensitivity of 91.3% and specificity of 73.9%, suggesting performance may drop across sites and datasets.

• The tool is best viewed as decision support, not a replacement for histopathology or clinical judgment.

• A custom graphical user interface and Grad-CAM overlays suggest the model could be deployed in a real-time, interpretable workflow during surgery.

Insights

• The investigators addressed marked class imbalance with focal loss, label smoothing, dropout regularisation, adaptive class weighting, and weighted sampling.

• Calibration was reasonable, with Brier scores of 0.16 internally and 0.20 externally, which matters if predicted probabilities are used in workflow decisions.

• The dataset was small and image-based: 275 images total, including 37 tumour images, so broader validation in larger and more diverse cohorts is still needed.

The Bottom Line

• This study suggests deep learning can help interpret prostatectomy FCM images quickly enough for intraoperative use, with promising discrimination and explainability.

• For clinicians, the near-term value is likely faster margin triage and reduced dependence on on-site pathology, pending larger prospective validation.