Introduction
Diabetic foot ulcers (DFUs) represent a significant clinical challenge, often leading to severe morbidity and lower-limb amputations. Accurate wound segmentation is critical for monitoring healing progress, yet manual annotation by clinicians is subjective and labor-intensive. While deep learning models, such as U-Net, have shown promise in medical image segmentation, they typically require large volumes of annotated data. In the medical domain, expert-labeled datasets are often scarce and expensive to produce. Furthermore, standard transfer learning techniques frequently struggle due to the significant domain gap between general skin lesion datasets and the complex, irregular appearance of foot ulcers.
The TransMix Framework
To address these challenges, we present TransMix, a two-stage cross-domain augmentation strategy designed to enhance segmentation performance in data-scarce environments. TransMix bridges the gap between large-scale dermatology datasets and small-scale ulcer datasets through two primary components: Augmented Global Pre-training (AGP) and Localized CutMix Fine-tuning (LCF).
Augmented Global Pre-training (AGP)
The first stage involves pre-training a U-Net model (with a ResNet-50 backbone) on the HAM10000 dataset, which contains over 10,000 skin lesion images. To improve robustness against the varied conditions of real-world DFU photography, AGP employs intense global transformations. These include perspective shifts, optical distortions, and significant color/brightness adjustments, teaching the model to recognize skin features across diverse lighting and camera angles.
Localized CutMix Fine-tuning (LCF)
During the fine-tuning phase on DFU data, TransMix utilizes a refined "CutMix" approach. Unlike traditional CutMix, which swaps random rectangular patches, LCF focuses on biological realism. It extracts actual wound candidates from the training set, removes their backgrounds to eliminate artificial edges, and pastes them onto target images using a "location-aware" strategy. This ensures the synthetic training samples better mimic the morphological characteristics of real ulcers.
Experimental Setup
The framework was evaluated on the FUSeg benchmark, a real-world foot ulcer dataset. To simulate extreme data scarcity, experiments were conducted using only 40 (5%) and 81 (10%) annotated images for training. The performance was compared against a baseline U-Net and several ablation settings.
Results and Discussion
TransMix demonstrates substantial improvements, particularly when training data is most limited. As shown in the performance analysis, the integration of AGP and LCF allows the model to significantly outperform the baseline U-Net.
Key findings include:
- Performance Gains: With only 40 training images, TransMix achieved a Dice score of 85.26%, a 10.43% improvement over the baseline U-Net (74.83%). With 81 images, the score reached 86.65%.
- Addressing the Domain Gap: Simple pre-training on skin lesions without AGP actually hindered performance, confirming that the domain gap between "clean" lesions and "messy" ulcers requires specific augmentation strategies to bridge.
- Synergy of Global and Local Augmentation: AGP addresses global variability (image-wide conditions), while LCF targets local variability (wound tissue appearance). The combination of both is essential for maximizing model diversity and accuracy.
Conclusion
The TransMix framework effectively mitigates the data scarcity problem in foot ulcer segmentation by leveraging cross-domain knowledge and realistic data augmentation. By achieving high accuracy with as few as 40 annotated images, TransMix reduces the annotation burden on medical experts. Future work will explore few-shot and zero-shot learning paradigms to further minimize the reliance on labeled data in clinical settings.
Summary updated