Can surgical computer vision benefit from large-scale visual foundation models?

Abstract

We investigate whether foundation models pretrained on diverse visual data could be beneficial to surgical computer vision. We use instrument and uterus segmentation in mini-invasive procedures as benchmarks. We propose multiple supervised, unsupervised and few-shot supervised adaptations of foundation models, including two novel adaptation methods. We use DINOv1, DINOv2, DINOv2 with registers, and SAM backbones, with the ART-Net surgical instrument and the SurgAI3.8K uterus segmentation datasets. We investigate five approaches: DINO unsupervised, few-shot learning with a linear decoder, supervised learning with the proposed DINO-UNet adaptation, DPT with DINO encoder, and unsupervised learning with the proposed SAM adaptation. We evaluate 17 models for instrument segmentation and 7 models for uterus segmentation and compare to existing ad hoc models for the tasks at hand. We show that the linear decoder can be learned with few shots. The unsupervised and linear decoder methods obtain slightly subpar results but could be considered useful in data scarcity settings. The unsupervised SAM model produces finer edges but has inconsistent outputs. However, DPT and DINO-UNet obtain strikingly good results, defining a new state of the art by outperforming the previous-best by 5.6 and 4.1 pp for instrument and 4.4 and 1.5 pp for uterus segmentation. Both methods obtain semantic and spatial precision, accurately segmenting intricate details. Our results show the huge potential of using DINO and SAM for surgical computer vision, indicating a promising role for visual foundation models in medical image analysis, particularly in scenarios with limited or complex data.