Related Experiment Video
Updated: Sep 13, 2026
Multispectral Real-time Fluorescence Imaging for Intraoperative Detection of the Sentinel Lymph Node in Gynecologic Oncology
Published on: October 20, 2010
GENERATIVE AI FOR HIRSCHSPRUNG DISEASE: CAN SYNTHETIC FLUORESCENCE CONFOCAL MICROSCOPY IMAGES ENHANCE INTRAOPERATIVE
Donatella Di Fabrizio1, Agnese Sbrollini2, Edoardo Bindi3
1Pediatric Surgery Unit, Salesi Children's Hospital, Polytechnic University of Marche, Ancona, Italy.
Background:
Accurate identification of ganglionated bowel is essential during laparoscopic pull-through for Hirschsprung Disease (HD), yet intraoperative biopsy interpretation is time-sensitive and operator-dependent. Fluorescence confocal microscopy (FCM) provides rapid imaging of fresh tissue, and deep-learning (DL) has the potential to extract diagnostic patterns from these images automatically. However, DL development is limited by HD rarity and images scarcity. Generative-AI may address this gap by synthesizing realistic images to expand underrepresented data, improving model robustness. This study aims to evaluate whether conditional generative adversarial network (cGAN)-synthetic FCM images can improve DL performance for discriminating ganglionic cells.
Methods:
Between November 2024 and November 2025, intraoperative FCM was routinely performed during laparoscopic pull-through for HD. High-resolution images were annotated to identify ganglionic and aganglionic regions, subdivided into 256×256-pixel tiles (resolution 0.5x0.5μm/pixel), and grayscale (128 levels) converted. Three convolutional neural networks (CNN), having the same architecture, were trained on different datasets: CNN-1 considered only real tiles (461 ganglionic and 1374 aganglionic), CNN-2 considered the same real tiles (461 ganglionic and 1374 aganglionic) and 913 conditional-GAN simulated ganglionic tiles, and CNN-3 considered the same real tiles (461 ganglionic and 1374 aganglionic) and 913 data augmented ganglionic tiles. CNNs were validated in the same independent testing dataset composed of real 590 aganglionic and 198 ganglionic tiles.
Results:
CNN-1, CNN-2 and CNN-3 identified ganglionic tiles with 75%, 84% and 80% accuracy, 78%, 91% and 79% sensitivity, 73%, 78% and 75% specificity, respectively. CNN-2 achieved a significantly higher ROC-AUC than CNN-1 (0.94 vs 0.84 and 0.80; P<0.01).
Conclusion:
Synthetic augmentation with cGAN-generated ganglionic tiles significantly improved CNN discrimination of ganglionic bowel on real FCM images, supporting generative-AI as a scalable strategy to strengthen DL for discriminating ganglionic bowel in HD.
