Interpreting Venous and Arterial Ulcer Images Through the Grad-CAM Lens: Insights and Implications in CNN-Based Wound Image Classification

Hannah Neuwieser, Naga Venkata Sai Jitin Jami, Robert Johannes Meier, Gregor Liebsch, Oliver Felthaus, Silvan Klein, Stephan Schreml, Mark Berneburg, Lukas Prantl, Heike Leutheuser, Sally KempaDiagnostics 2025 · OpenReview

TL;DR

This paper explores the use of deep learning models for classifying venous and arterial leg ulcers, achieving high accuracy. By employing Grad-CAM, the researchers not only demonstrate the models' diagnostic capabilities but also gain insights into how they make decisions, enhancing trust and interpretability for potential clinical adoption.

What Problem Does This Paper Address?

Chronic lower extremity ulcers, specifically arterial and venous types, pose a significant clinical and economic challenge. Differentiating between these two types is crucial for effective treatment, as misdiagnosis can lead to delayed or inappropriate care, potentially resulting in severe complications like limb amputation. While traditional diagnostic methods exist (e.g., Doppler ultrasound), they can be time-consuming, subjective, and not universally accessible. This paper addresses the need for a faster, more objective, and potentially more accessible diagnostic aid by investigating the efficacy of Convolutional Neural Networks (CNNs) for classifying these ulcers from images. Furthermore, it tackles the "black box" problem of deep learning by using explainable AI (XAI) techniques to understand the models' decision-making processes, which is vital for clinical acceptance.

Methods

The study evaluates five state-of-the-art CNN architectures for binary classification of arterial versus venous ulcers: ResNet50, ResNeXt50, ConvNeXt (Tiny variant), EfficientNetB2, and EfficientNetV2. These models were chosen for their proven performance in medical imaging and architectural diversity.

Key methodological aspects include:

  • Data Augmentation: A weak augmentation strategy (horizontal/vertical flips, random cropping) was applied dynamically during training to increase dataset variety without creating new entries, aiming to improve model robustness.
  • Training Protocol: All models were trained using a unified protocol with a batch size of 32, an initial learning rate of 1x10⁻⁴, and trained for 50 epochs.
  • Performance Evaluation: Standard classification metrics were used: Accuracy, Precision, Recall, F1-Score, and ROC-AUC.
  • Interpretability: Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize the image regions that most influenced the models' predictions. This technique highlights areas of high importance (red/yellow) versus low importance (blue/purple) for classification.

Datasets

The study utilized a retrospective dataset of 607 wound images collected from the University Hospital Regensburg, Germany. This dataset comprised 198 images of arterial ulcers and 409 images of venous ulcers. The images were originally of varying sizes and were manually cropped to focus on the wound and a small surrounding margin. No new benchmark dataset was introduced.

Key Insights and Ideas

  • High Classification Performance: Advanced CNNs can achieve very high accuracy in distinguishing between arterial and venous ulcers, suggesting their potential as a valuable diagnostic tool.
  • EfficientNet Family's Strength: EfficientNet models (B2 and V2) demonstrated strong performance, particularly in terms of precision, and showed a more "dynamic" focus in Grad-CAM visualizations, attending to various parts of the wound and surrounding tissue.
  • Model-Specific Attention: Different CNN architectures focus on distinct image features. For instance, EfficientNetV2 seemed to consider ulcer borders and surrounding skin, while ResNeXt50 tended to focus more on the central wound bed. This highlights how architectural choices can influence what features a model prioritizes.
  • Interpretability Builds Trust: Grad-CAM visualizations provide crucial insights into the models' decision-making, making their predictions more transparent and understandable. This is essential for gaining clinician trust and facilitating the integration of AI into clinical workflows.
  • Lightweight Models are Viable: The study suggests that even relatively lightweight models can achieve high performance and provide interpretable results, making them more practical for deployment.

Key Results

  • Accuracy: Models achieved high accuracies, ranging from 72% (ConvNeXt) to 98% (ResNeXt50).
  • Top Performers: ResNeXt50 achieved the highest macro-average accuracy (97.85%) and ROC-AUC (0.9966). EfficientNetV2 also performed very well with high precision.
  • ConvNeXt's Underperformance: The ConvNeXt (Tiny) model showed significantly lower performance (72.04% accuracy) compared to the others.
  • Grad-CAM Insights: Visualizations showed that EfficientNet models exhibited a more context-aware focus, examining different parts of the image, while ResNeXt50 tended to fixate on the central wound bed. This dynamic attention in EfficientNet models aligns with their high precision and suggests better adaptability to wound variations.

Limitations and Future Work

  • Single-Center Dataset: The data was collected from a single hospital, limiting generalizability. Future work should involve larger, multicenter datasets with diverse imaging conditions.
  • Binary Classification: The study focused only on distinguishing between arterial and venous ulcers. Future research should expand to include mixed arterial-venous ulcers, diabetic foot ulcers, and other etiologies.
  • Image-Only Approach: The models relied solely on photographs, excluding crucial clinical data like patient history, vascular imaging, and physical examination findings. Incorporating these multimodal data streams could significantly enhance diagnostic accuracy.
  • Manual Cropping: Images were manually cropped, which is not scalable for real-world applications. Future work should explore automatic wound detection and cropping algorithms.
  • Qualitative XAI: Grad-CAM provides qualitative insights. Future work could incorporate quantitative XAI methods like SHAP for more detailed per-pixel importance scores.
  • Prospective Validation: The study was retrospective. Prospective external validation is needed to confirm the models' real-world efficacy.

Summary updated

Abstract

Background/Objectives: Chronic wounds of the lower extremities, particularly arterial and venous ulcers, represent a significant and costly challenge in medical care. To assist in differential diagnosis, we aim to evaluate various advanced deep-learning models for classifying arterial and venous ulcers and visualize their decision-making processes. Methods: A retrospective dataset of 607 images (198 arterial and 409 venous ulcers) was used to train five convolutional neural networks: ResNet50, ResNeXt50, ConvNeXt, EfficientNetB2, and EfficientNetV2. Model performance was assessed using accuracy, precision, recall, F1-score, and ROC-AUC. Grad-CAM was applied to visualize image regions contributing to classification decisions. Results: The models demonstrated high classification performance, with accuracy ranging from 72% (ConvNeXt) to 98% (ResNeXt50). Precision and recall values indicated strong discrimination between arterial and venous ulcers, with EfficientNetV2 achieving the highest precision. Conclusions: AI-assisted classification of venous and arterial ulcers offers a valuable method for enhancing diagnostic efficiency.
BibTeX
@inproceedings{neuwieser2025interpreting,
  author = {Hannah Neuwieser and Naga Venkata Sai Jitin Jami and Robert Johannes Meier and Gregor Liebsch and Oliver Felthaus and Silvan Klein and Stephan Schreml and Mark Berneburg and Lukas Prantl and Heike Leutheuser and Sally Kempa},
  title = {{Interpreting Venous and Arterial Ulcer Images Through the Grad-CAM Lens: Insights and Implications in CNN-Based Wound Image Classification}},
  booktitle = {Diagnostics},
  year = {2025},
  url = {https://openreview.net/forum?id=WYUJMKl1sS},
}
Figure