Article intro - The LEMON dataset and surgical FM

LEMON: A Large Endoscopic MONocular Dataset and Foundation Model for Perception in Surgical Settings Chengan Che, Chao Wang, Tom Vercauteren, Sophia Tsoka, Luis C. Garcia-Peraza-Herrera presented by  the VISURG lab at King's College London, accepted to #CVPR2026:

 "While existing open-access datasets have laid an amazing groundwork for major breakthroughs, the shift toward highly generalizable foundation models requires an entirely new scale of massive and diverse data.

𝗧𝗵𝗲 𝗟𝗘𝗠𝗢𝗡 𝗗𝗮𝘁𝗮𝘀𝗲𝘁: We compiled an extensive collection of over 4,000 high-resolution surgical videos spanning 35 distinct procedure types (both robotic and traditional). That is 𝟵𝟯𝟴 𝗵𝗼𝘂𝗿𝘀 (𝟴𝟱 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝗳𝗿𝗮𝗺𝗲𝘀) of high-quality footage—a massive leap in size and scope compared to existing alternatives.

𝗟𝗲𝗺𝗼𝗻𝗙𝗠 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹: To prove the effectiveness of this diverse data, we built LemonFM. It is pretrained on the LEMON dataset using a novel self-supervised augmented knowledge distillation approach.

𝗦𝘁𝗮𝘁𝗲-𝗼𝗳-𝘁𝗵𝗲-𝗔𝗿𝘁 𝗥𝗲𝘀𝘂𝗹𝘁𝘀: LemonFM consistently outperforms existing surgical foundation models across 𝗳𝗼𝘂𝗿 𝗱𝗼𝘄𝗻𝘀𝘁𝗿𝗲𝗮𝗺 𝘁𝗮𝘀𝗸𝘀 and 𝘀𝗶𝘅 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀, achieving significant gains in surgical phase recognition and beyond.

𝗨𝗻𝗺𝗮𝘁𝗰𝗵𝗲𝗱 𝗗𝗮𝘁𝗮 𝗘𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆: Getting labeled surgical data is expensive. LemonFM is so robust that even when fine-tuned on just 50% of the labeled data, it still outperforms existing surgical foundation models trained on 100% of the data!

𝗥𝗲𝗮𝗹-𝗧𝗶𝗺𝗲 𝗥𝗲𝗮𝗱𝘆: In surgery, latency matters. We optimized LemonFM for clinical viability, achieving an inference time of just 7.8 ms per frame —making it highly efficient and ready for real-time surgical applications.

Huge shoutout to the incredible team behind this: Chengan Che, Chao Wang, Tom Vercauteren, and Sophia Tsoka!

We built LEMON to be a comprehensive resource for the community, and we can't wait to see what you build with it."

𝗥𝗲𝗮𝗱 𝘁𝗵𝗲 𝗳𝘂𝗹𝗹 𝗽𝗮𝗽𝗲𝗿 𝗵𝗲𝗿𝗲: https://lnkd.in/d2a3xXCa 
𝗖𝗼𝗱𝗲 & 𝗗𝗮𝘁𝗮𝘀𝗲𝘁: https://lnkd.in/dPeNX4qg 
𝗟𝗲𝗮𝗿𝗻 𝗺𝗼𝗿𝗲 𝗮𝗯𝗼𝘂𝘁 𝗼𝘂𝗿 𝗹𝗮𝗯: visurg.ai 

Source: LinkedIn