Interpretable Deep Learning Applied to Fluorescence Confocal Microscopy for Intraoperative Tissue Assessment in...

In pediatric surgical oncology, intraoperative tissue assessment is limited by small specimen size and the absence of real-time histopathology. Ex vivo fluorescence confocal microscopy (FCM) provides rapid histology like imaging of fresh tissue, yet interpretation relies on expert visual assessment. We tested the hypothesis that deep learning applied directly to ex vivo FCM images enables automated, accurate, and spatially interpretable tumor detection in pediatric specimens.

FCM images were prospectively acquired during routine clinical activity from 42 children, yielding 141 surgical and biopsy specimens. For this analysis, anonymized mosaics were retrospectively collected and manually annotated as malignant, benign, or healthy tissue by expert, then decomposed into 256 × 256 pixel tiles at 0.5 µm per pixel. A convolutional neural network was trained, validated, and tested on stratified datasets. Performance metrics were calculated at the tile level using accuracy, sensitivity, specificity, and F1 score. Tile level activation heatmaps were generated to localize regions influencing predictions.

Overall, 243,659 tiles were analyzed, including 124,348 malignant, 48,441 benign, and 70,870 healthy tissue tiles. The independent testing dataset comprised 84,751 tiles. Malignant tissue was identified with 91.35% accuracy, 95.88% sensitivity, 86.28% specificity, and an F1 score of 92.13%. Healthy tissue achieved 94.41% accuracy and 95.58% specificity. Benign tissue showed 90.14% accuracy and 91.78% specificity. Heatmaps consistently highlighted at a cellular level the architectural distortion in malignant tiles.

This work introduces the first pediatric integration of deep learning with ex vivo FCM, demonstrating the feasibility of rapid and interpretable tissue classification and providing the basis for future prospective intraoperative validation.