Source-linked AI summary
BruNet: A Cross-Domain Transfer Framework for Bruise Segmentation
Qiming Wang, Richard J. Motley, Ebube E. Obi, Xianfang Sun, Paul L. Rosin
TL;DR
Bruise segmentation is hindered by scarce annotations, variable appearance, and diffuse boundaries, while prior work has focused mainly on detection or classification. BruNet combines a ViT encoder with a SAM decoder, transfers from HAM10000 to bruise images without fine-tuning, and outperforms the evaluated baselines under domain shift.
Problem
Bruise segmentation lacks established pixel-level methods and is constrained by scarce annotations, diffuse boundaries, and variable appearance.
Method
BruNet combines a ViT-based visual encoder with a SAM mask decoder, trains on HAM10000 skin-lesion masks, and evaluates on bruise images without task-specific fine-tuning.
Results
BruNet consistently outperforms CNN baselines, prompt-based SAM variants, and other evaluated models across Dice and IoU, with BruNet-LingBot reaching Dice 0.8674 and IoU 0.7897.
Takeaways & Limitations
The findings support cross-domain foundation-model transfer for bruise segmentation in visually ambiguous, under-annotated clinical targets.
Takeaways & Limitations
The dual-region evaluation protocol is a pragmatic surrogate for multi-annotator evaluation rather than a substitute for independent inter-observer variability.
Abstract
from arXiv · showhide
Segmenting bruises is a challenging task in medical imaging due to limited data and annotations, diffuse boundaries, and highly variable appearance. In this work, we propose BruNet, a segmentation framework that combines a ViT-based visual encoder (a self-supervised DINOv3 or a pretrained LingBot-Vision backbone) with a SAM-based mask decoder. BruNet is trained on the HAM10000 skin lesion dataset and evaluated on a separate bruise dataset without additional fine-tuning. Although a small number of prior studies have explored machine learning and computer vision for bruise analysis, existing work has primarily focused on detection, classification, or colour analysis rather than pixel-level localisation. To the best of our knowledge, this is the first study to address automatic bruise segmentation. Our results show that BruNet outperforms CNN-based models, state-of-the-art segmentation models, ChatGPT-4o/5-assisted SAM2 zero-shot baselines, and the medical-oriented MedSAM model, demonstrating strong cross-domain generalisation to bruise segmentation.
1. Introduction
Bruise segmentation is difficult because bruises vary substantially in appearance, have diffuse boundaries, and lack publicly available annotated datasets. BruNet addresses this gap with cross-domain transfer from skin-lesion segmentation to bruise segmentation.
- Motivation: Bruise analysis is underexplored despite clinical and forensic importance, partly because bruises vary across individuals and over time and are rarely documented with pixel-level annotations.Bruises may reflect haemostatic, vascular, or systemic disease and can provide evidence in forensic and legal contexts.
- Proposed approach: BruNet combines a ViT-based visual backbone with a SAM decoder, trains on HAM10000 skin-lesion masks, and evaluates on a separate bruise dataset.The framework is designed to assess cross-domain generalisation without relying on bruise-specific training data.
- Results: BruNet results show robustness under domain shift against CNN, transformer, prompt-driven SAM2, and medical-oriented baselines.The comparison includes MedSAM with MedGemma as a prompter.
- Contributions: The study introduces the first deep learning framework for bruise segmentation together with an expert-annotated bruise dataset and a dual-region protocol for uncertain boundaries.The protocol distinguishes high-confidence bruise cores from uncertain boundary regions.
2. Related Works
Prior medical-imaging methods often require extensive annotations, while bruise analysis remains focused mainly on detection and classification rather than pixel-level localisation. Bruises are especially difficult because their subsurface colour changes produce weak, diffuse boundaries unlike many visible skin abnormalities.
- Vision-transformer pretraining: Self-supervised DINOv3 and boundary-aware LingBot-Vision provide transferable representations, but their effectiveness for diffuse, low-contrast bruises remains untested.Both backbones were benchmarked on natural images with better-defined object boundaries.
- Foundation models: SAM offers promptable zero-shot segmentation, whereas MedSAM improves medical boundary fidelity but still inherits assumptions about objectness and edge contrast.The passage frames these assumptions as a limitation for medical boundaries that differ from common objects.
- Distinctive bruise challenge: Bruises differ from lesions, wounds, and burns because subsurface, low-contrast colour changes create weaker surface structure and greater boundary ambiguity.Lighting, skin tone, and skin thickness further contribute to bruise variability.
- Medical segmentation: CNN and ViT segmentation models can perform strongly but often lack generalisability and require large, high-quality annotated datasets.These requirements are particularly problematic for data-scarce targets such as bruises.
- Bruise-analysis gap: Existing bruise-analysis approaches primarily address detection or classification, leaving precise pixel-level segmentation comparatively underexplored.This gap concerns automatic localisation of bruise regions rather than broader bruise recognition tasks.
3. Methods
BruNet transfers visual representations learned from HAM10000 skin lesions to bruise segmentation through a shared SAM decoder. Its adapter maps different ViT token grids to a common dense representation, while augmentation and combined Dice–BCE training support segmentation under domain shift.
- Architecture: BruNet uses LingBot-Vision or DINOv3 as a visual encoder, a convolutional upsampling adapter, and the SAM prompt encoder and mask decoder.The two backbone options use masked-boundary or self-supervised pretraining, respectively.
- Architecture: The adapter maps LingBot-Vision’s 32 × 32 or DINOv3’s 14 × 14 token grid to a shared 56 × 56 × 256 embedding for the unchanged SAM decoder.It reshapes tokens, projects channels, upsamples spatially, and applies two 3 × 3 convolutional refinement blocks.
- Data and transfer setting: Training uses HAM10000 skin-lesion masks while reserving 86 expert-annotated bruise images exclusively for zero-shot evaluation.The setup treats bruise segmentation as transfer under scarce pixel-level annotations.
- Data and transfer setting: HAM10000 supplies diverse lesion morphologies whose irregular shapes, diffuse boundaries, colour variation, and heterogeneous textures provide transferable structural priors.The stated rationale is that these priors are not tied to a specific semantic class.
- Training strategy: MSRCP Retinex augmentation is applied to encourage invariance to illumination and contrast changes in dermoscopic images.The augmentation addresses appearance changes caused by differing illumination conditions during HAM10000 collection.
- Training strategy: The total training loss combines Dice loss and binary cross-entropy, while the two variants use unfrozen visual backbones and LoRA-adapted SAM decoders.The LoRA configuration uses rank r = 8 and scaling factor α = 8.
4. Experiments
Experiments evaluate BruNet against conventional, transformer, and prompt-based segmentation baselines on an 86-image bruise dataset using uncertainty-aware annotations and standard overlap metrics. BruNet variants achieve the strongest overall performance without bruise-specific training, while statistical tests confirm their advantage over non-BruNet baselines.
- Evaluation Setup: The evaluation used expert-reviewed dual-region annotations separating a high-confidence inner bruise region from a broader uncertain boundary region.The protocol was designed because multiple independent expert annotations were impractical for bruise images.
- Evaluation Setup: The dual-region protocol is a pragmatic surrogate rather than a substitute for multi-expert evaluation because candidate masks depend on one expert annotation and model-generated variants.A proper multi-expert annotation study remains future work.
- Results: BruNet-LingBot achieved the highest Dice (0.8674), IoU (0.7897), and accuracy (0.9556), while BruNet-DINOv3 achieved the highest recall (0.8491).Both variants were trained only on HAM10000 skin lesion masks and evaluated on bruises without bruise-specific fine-tuning.
- Results: OneFormer was the strongest conventional baseline with Dice 0.7097 and IoU 0.6352, but its precision of 0.7063 indicated inclusion of non-bruise regions.U-Net showed stronger recall than precision and produced a lower Dice score of 0.5198, consistent with over-segmentation.
- Results: Prompted SAM2 and MedSAM pipelines generally produced conservative or incomplete masks, with GPT-5 + SAM2 reaching precision 0.9467 but recall only 0.2924.MedGemma + MedSAM was the strongest LLM-prompted pipeline at Dice 0.5157 and IoU 0.3911, remaining below OneFormer, SAM3, and BruNet.
- Results: SAM3 achieved baseline precision 0.9372 and accuracy 0.9208, but its recall of 0.5888 reduced Dice to 0.7053 and IoU to 0.5875.SAM3 returned no retained bruise instance for five of the 86 test images.
- Statistical Analysis: BruNet-DINOv3 outperformed all eight non-BruNet baselines for Dice and IoU after Holm–Bonferroni correction, including OneFormer with adjusted p-values of 0.0044 and 0.0059.Comparisons used paired per-image scores from the same 86 test images.
- Statistical Analysis: The two BruNet backbones had comparable image-level Dice and IoU distributions, although LingBot achieved higher arithmetic means.Neither the Dice nor IoU difference between LingBot and DINOv3 was statistically significant.
5. Ablation Study
The ablations show that both the SAM decoder and intermediate feature upsampling materially improve BruNet’s segmentation quality. Removing either component lowers overlap metrics, while decoder replacement increases recall through over-segmentation.
- Decoder and upsampling ablations: Dice fell from 0.867 to 0.848 and IoU from 0.790 to 0.764 when the SAM decoder was replaced by a convolutional head.The convolutional head achieved higher recall but lower precision, indicating over-segmentation rather than better delineation.
- Decoder and upsampling ablations: Removing intermediate 56 × 56 upsampling reduced Dice to 0.839 and IoU to 0.752, producing more fragmented masks and lower recall.The native 32 × 32 feature grid retained high precision but captured mainly high-confidence bruise cores.
- Decoder and upsampling ablations: The ablations support intermediate upsampling and the SAM decoder as contributors to spatial continuity and complete bruise-region recovery.Both components were evaluated under the same training data and optimisation settings.
6. Conclusion
BruNet transfers segmentation knowledge from HAM10000 skin lesions to bruise images without task-specific fine-tuning, achieving strong cross-domain performance. The study introduces uncertainty-aware evaluation while identifying future directions for stronger foundation models and broader bruise assessment.
- Replacing the SAM decoder or removing intermediate feature upsampling reduces segmentation performance, indicating both components contribute to the final design.
- BruNet demonstrates strong generalisation to out-of-distribution bruise images without task-specific fine-tuning after training on HAM10000.
- BruNet consistently outperforms CNN baselines and prompt-based SAM variants across Dice and IoU under the dual-region evaluation.
- The dual-region evaluation reflects ambiguity in bruise boundaries, while the 86-image dataset provides expert-reviewed annotations for assessment.
- 6.1.1. Integration of foundation models: Future work could replace the ViT backbone with domain-specific foundation models and integrate bruise-age or severity prediction beyond localisation.