# TransUNet-GradCAM: A Hybrid Transformer-U-Net with Self-Attention and Explainable Visualizations for Foot Ulcer Segmentation

arXiv: https://arxiv.org/abs/2508.03758 · HTML: https://mlsanity.com/paper/2508.03758/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 15 of its 53 references that are in the map.
- First posted on arXiv 4 Aug 2025.

## Builds on (most cited first, 15 of 15)

- [Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization](https://mlsanity.com/paper/1610.02391/) (2016)
- [Quantifying Attention Flow in Transformers](https://mlsanity.com/paper/2005.00928/) (2020)
- [Albumentations: fast and flexible image augmentations](https://mlsanity.com/paper/1809.06839/) (2018)
- [Fully Automatic Wound Segmentation with Deep Convolutional Neural Networks](https://mlsanity.com/paper/2010.05855/) (2020)
- [Few-shot Adaptation of Medical Vision-Language Models](https://mlsanity.com/paper/2409.03868/) (2024)
- [FUSeg: The Foot Ulcer Segmentation Challenge](https://mlsanity.com/paper/2201.00414/) (2022)
- [Deep Learning in Diabetic Foot Ulcers Detection: A Comprehensive Evaluation](https://mlsanity.com/paper/2010.03341/) (2020)
- [Scheduling Techniques for Liver Segmentation: ReduceLRonPlateau Vs OneCycleLR](https://mlsanity.com/paper/2202.06373/) (2022)
- [FUSegNet: A Deep Convolutional Neural Network for Foot Ulcer Segmentation](https://mlsanity.com/paper/2305.02961/) (2023)
- [Weighted Intersection over Union (wIoU) for Evaluating Image Segmentation](https://mlsanity.com/paper/2107.09858/) (2021)
- [Automatic Foot Ulcer Segmentation Using an Ensemble of Convolutional Neural Networks](https://mlsanity.com/paper/2109.01408/) (2021)
- [Wound Tissue Segmentation in Diabetic Foot Ulcer Images Using Deep Learning: A Pilot Study](https://mlsanity.com/paper/2406.16012/) (2024)
- [Efficient In-Context Medical Segmentation with Meta-driven Visual Prompt Selection](https://mlsanity.com/paper/2407.11188/) (2024)
- [Understanding implementation pitfalls of distance-based metrics for image segmentation](https://mlsanity.com/paper/2410.02630/) (2024)
- [Deploying and Evaluating Multiple Deep Learning Models on Edge Devices for Diabetic Retinopathy Detection](https://mlsanity.com/paper/2506.14834/) (2025)

## Built on by (most cited first, 2 of 2)

- [Improved Segmentation of Polyps and Visual Explainability Analysis](https://mlsanity.com/paper/2509.18159/) (2025)
- [SSPFormer: Self-Supervised Pretrained Transformer for MRI Images](https://mlsanity.com/paper/2601.12747/) (2026)

## Abstract (arXiv)

Automated segmentation of diabetic foot ulcers (DFUs) supports clinical diagnosis, treatment planning, and wound monitoring, but remains challenging because of heterogeneous appearance, irregular morphology, and cluttered backgrounds in clinical photographs. We evaluate a hybrid ViT-bottleneck U-Net that combines a convolutional encoder-decoder with a Transformer bottleneck and attention-gated skip connections. We emphasise rigorous validation and explainability rather than architectural novelty. The model was trained on the public Foot Ulcer Segmentation Challenge (FUSeg) dataset using a hybrid Dice and cross-entropy loss. Results are reported over five seeds as mean +/- 95% confidence interval at a fixed threshold. On the internal validation set, the model achieved a Dice of 0.8035 +/- 0.0053 and IoU of 0.7149 +/- 0.0073 (HD95 = 19.74 px, ASSD = 6.12 px). Ablation showed that only the hybrid loss significantly changed Dice (-0.038, p < 0.001), while the Transformer bottleneck, attention gates, and augmentation had small, non-significant in-domain effects. External validation without retraining achieved a Dice of 0.7460 on the AZH Wound Care Center cohort (n = 278), retaining about 92% of internal Dice. A small Medetec subset (n = 8) was used only for qualitative assessment, indicating partial rather than robust generalisation under domain shift. Explainability analysis found Grad-CAM more wound-localised (energy-in-mask 0.871 vs. 0.102), while attention rollout was significantly more faithful (p = 0.038, n = 200). Predicted and expert wound areas showed strong agreement (Pearson r = 0.944), with a lightweight model of 8.79 M parameters.