Adapted DINOv2 generalises better than task-specific medical segmentation nets
Original title: Medical Image Segmentation: Foundation Model Adaptation versus Task-Specific Training
Only a small part of DINOv2 was adapted; in-distribution U-Nets stayed sharper and caught small lesions more often.
Researchers adapted DINOv2 from bias-term tuning through full fine-tuning and compared it with U-Nets and ResU-Nets trained from scratch. Experiments covered colonoscopy, breast ultrasound and dermoscopy, with held-out source data and independent datasets from other institutions. The foundation model generalised better on every external set, while task-specific models stayed competitive when training data were abundant. Required adaptation capacity depended on the imaging domain, and a widely used polyp test set duplicated images from another benchmark.
Unconfirmed: Generalisation and duplication findings are reported by the paper's authors.
- The adapted model more often missed small lesions inside the training distribution.
- Required capacity tracked the imaging domain rather than the adaptation method.
- A widely used polyp test set duplicated images from another benchmark.
From the original article
1 min read at arxiv.orgFoundation models pretrained on large natural-image collections are increasingly adapted for medical image segmentation, in the expectation that they will replace models trained for each task. Most comparisons, however, use data from the training distribution and a single task-specific baseline whose configuration is rarely justified, and therefore provide limited evidence of performance at new clinical sites.
By Emre Kutlusan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.