AI-SLN: AI-based analysis of ICG fluorescence for sentinel lymph node assessment in endometrial cancer surgery
Call: Fondazione Cassa di Risparmio in Bologna, Bando Ricerca Scientifica e Alta Tecnologia 2024
Period: 2024 to 2026
Principal Investigator: Prof. Gianluca Moro (DISI, University of Bologna)
Clinical partner: IRCCS AOU di Bologna, Policlinico di Sant’Orsola, Gynaecology and Human Reproduction Physiopathology (Prof. Renato Seracchioli, Dr. Diego Raimondo, Dr. Alberto Aguzzi)
Project website: disi-unibo-nlp.github.io/ai-sln
Endometrial cancer is the most common gynaecological malignancy in Europe, and its treatment depends on whether the tumour has spread to the pelvic lymph nodes. Instead of removing the whole nodal chain, surgeons target the sentinel lymph node, marked with the fluorescent dye indocyanine green (ICG) and identified on a near-infrared laparoscopic camera. The technique is the standard of care, but it fails on one side in about 23.5% of patients, often because the removed tissue turns out to contain no lymph node at all (an “empty node packet”).
AI-SLN is a pilot study exploring whether computer vision and vision-language models can read intra-operative ICG laparoscopic video to support the surgeon’s decisions. The project frames three clinical questions: detecting whether the fluorescent tissue actually contains a sentinel node, anticipating the biopsy outcome (metastatic or disease-free) directly from the intra-operative image, and estimating how many nodes the tissue holds before removal. In every case the goal is decision support at the surgeon’s side, never an autonomous decision.
Working with anonymised videos and histopathology outcomes from 40 patients, and with no foundation model available for this image modality, we compared medical foundation encoders and multimodal large language models, combined with fine-tuning and data augmentation for extreme data scarcity, entirely on local infrastructure. The best models localise the first appearance of the sentinel node in two out of three clips, separate metastatic from disease-free nodes well above chance (PR-AUC 0.71), and estimate the node count within one node in 71% of cases.
These results provide a reproducible experimental base and motivate large-scale data collection and dedicated foundation models for fluorescence-guided surgery. The full description of the clinical problem, tasks, methods and results, together with an interactive demo on de-identified examples, is available on the AI-SLN project website.
Funding
This project is supported by Fondazione CARISBO through the competitive call “Bando Ricerca Scientifica e Alta Tecnologia 2024”.