Self-supervised pretraining for dermoscopic lesion segmentation with limited labels: A controlled comparison of initialisation strategies
Year: 2026
Type: preprint
Abstract: Background Self-supervised learning can improve medical image segmentation when labelled data are scarce, but its practical benefit may depend heavily on the model architecture and initialisation strategy. This study evaluated the downstream utility of in-domain masked-reconstruction pretraining across strictly limited annotation budgets. Methods The International Skin Imaging Collaboration (ISIC) 2016 development set was partitioned into 765 images for self-supervised pretraining, 94 for fine-tuning and 41 for validation. Five nested annotation budgets (ranging from 12 to 94 images) across five training seeds were evaluated on the 379-image test set. A custom U-Net compared random and in-domain self-supervised initialisation. Concurrently, a ResNet-34 U-Net compared random, ImageNet, pretrained, pure self-supervised and sequential ImageNet-to-self-supervised initialisation pathways. Models trained with 94 labelled images were further evaluated without adaptation on ISIC 2017 to test cross-dataset robustness and a separate experiment examined an alternative restoration-based pretraining strategy. Results In the custom U-Net, self-supervised pretraining achieved higher mean Dice than random initialisation at all five budgets, including 0.664 versus 0.615 with 12 labels and 0.811 versus 0.768 with 94 labels. Seed-level 95% confidence intervals for paired differences excluded zero at 47 and 94 labels but not at the other budgets. Notably, with just 47 labels, the self-supervised model achieved a mean Dice of 0.789, surpassing the randomly initialised model trained on double the data (94 labels; mean Dice 0.768). Conversely, in the ResNet-34 architecture, ImageNet initialisation outperformed in-domain self-supervised pretraining at every budget, and appending self-supervised training after ImageNet actively reduced mean Dice at four of the five budgets. All six 94-label configurations experienced severe performance drops on ISIC 2017 relative to ISIC 2016. In the separate restoration experiment, mean Dice degraded from 0.686 under random initialisation to 0.638-0.642 following pretraining. Conclusions In-domain masked-reconstruction pretraining improved mean Dice relative to random initialisation in a custom U-Net from scratch but failed to provide a consistent advantage in a standard ResNet-34 U-Net, where baseline ImageNet initialisation performed best at every budget tier. Adding a self-supervised stage after ImageNet provided no measurable gain. These findings remain highly specific to the tested dermoscopic segmentation configurations and do not establish melanoma-specific diagnostic performance or clinical utility.
Source: medRxiv
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222042228