Work

2025Vision

AIMS: Surgical Phase Detection

Research intern, AIMS (AI in Medicine and Surgery)

I wanted a model that could name the phase of an endoscopic submucosal dissection from the video, with DINOv2, when the labels are sparse. The work with AIMS at Leeds, from March to November 2025, became the ISBI 2026 paper “Self-Supervised Vision Transformer for Surgical Phase Recognition in Endoscopic Submucosal Dissection” (DOI 10.1109/isbi61048.2026.11515812). The paper reports 89.5% accuracy on the patient set and 90.0% on the porcine set.

Endoscopic frames beside DINOv2 patch-norm and centre-similarity maps through an ESD sequence.
Fig. 1DINOv2 feature diagnostics across an endoscopic submucosal dissection sequence. Selected frames from one ESD sequence.
Stack
DINOv2, V-JEPA, Medical AI, Self-Supervised Learning
Links
Paper, Code

Also