Source-linked AI summary
Fully Convolutional Networks for Diabetic Foot Ulcer Segmentation
Manu Goyal, Neil D. Reeves, Satyan Rajbhandari, Jennifer Spragg, Moi Hoon Yap
TL;DR
DFU can lead to amputation, while existing assessment and segmentation approaches face clinical and technical limitations. The paper introduces a 705-image foot dataset with expert annotations and trains FCNs using two-tier transfer learning. Across 5-fold cross-validation, the models achieve Dice scores of 0.794 for ulcer, 0.851 for surrounding skin, and 0.899 for the combined region.
Problem
DFU assessment is clinically important because DFU can lead to amputation, while existing computer methods struggle with imaging variation, irregular contours, and surrounding-skin segmentation.
Method
The paper provides expert annotations for DFU and surrounding skin and trains fully convolutional networks using two-tier transfer learning from ImageNet and Pascal VOC.
Results
Dice scores were 0.794 for ulcer region, 0.851 for surrounding skin region, and 0.899 for the complete area, while healthy-foot specificity reached 1.0.
Takeaways & Limitations
The results demonstrate the potential of FCNs for automatically segmenting DFU and surrounding skin.
Abstract
from arXiv · showhide
Diabetic Foot Ulcer (DFU) is a major complication of Diabetes, which if not managed properly can lead to amputation. DFU can appear anywhere on the foot and can vary in size, colour, and contrast depending on various pathologies. Current clinical approaches to DFU treatment rely on patients and clinician vigilance, which has significant limitations such as the high cost involved in the diagnosis, treatment and lengthy care of the DFU. We introduce a dataset of 705 foot images. We provide the ground truth of ulcer region and the surrounding skin that is an important indicator for clinicians to assess the progress of ulcer. Then, we propose a two-tier transfer learning from bigger datasets to train the Fully Convolutional Networks (FCNs) to automatically segment the ulcer and surrounding skin. Using 5-fold cross-validation, the proposed two-tier transfer learning FCN Models achieve a Dice Similarity Coefficient of 0.794 ($\pm$0.104) for ulcer region, 0.851 ($\pm$0.148) for surrounding skin region, and 0.899 ($\pm$0.072) for the combination of both regions. This demonstrates the potential of FCNs in DFU segmentation, which can be further improved with a larger dataset.
I. INTRODUCTION
DFU assessment remains difficult because surrounding skin is clinically important, while existing computer-vision methods struggle with variable imaging conditions and irregular wound boundaries. The paper therefore proposes automated FCN-based segmentation of DFU and surrounding skin using a larger annotated dataset and two-tier transfer learning.
- DFU can lead to lower-limb amputation, making timely recognition and treatment clinically important.
- Surrounding skin indicates healing status and vulnerability to DFU extension, so it is an important assessment target.
- Existing computer methods use hand-crafted features, image processing, or manually tuned parameters that are sensitive to skin shades, illumination, resolution, and irregular contours.
- The paper proposes automated segmentation of DFU and surrounding skin using fully convolutional networks.
- The contributions include an annotated DFU dataset, separate surrounding-skin segmentation, and two-tier transfer learning from larger image datasets.
II. METHOD
The method section covers dataset preparation, expert annotation, conventional and deep-learning segmentation methods, and validation metrics.
- The methodology describes dataset preparation, expert labelling of DFU and surrounding skin, segmentation methods, and validation metrics.
A. DFU Dataset
The dataset contains DFU and healthy-foot images with expert-validated annotations, including separate labels for ulcer and surrounding skin regions. Images and labels are prepared for semantic segmentation using Pascal VOC format.
- The dataset includes 600 DFU images and 105 healthy foot images collected at Lancashire Teaching Hospitals.
- A podiatrist specializing in diabetic foot care created the annotations, which were validated by a diabetes consultant.
- Each DFU image labels background, surrounding skin, and DFU as separate classes in an 8-bit Pascal VOC-format image.
- The segmentation architecture uses convolutional and max-pooling layers for pixel-wise prediction.
B. Fully Convolutional Networks for DFU segmentation
Fully convolutional networks address the localization limitations of classification networks by producing pixel-wise predictions for image segmentation.
- Classification networks assign class probabilities but do not determine multiple objects’ positions within an image.
- Fully convolutional networks overcome this limitation by producing pixel-wise predictions for segmentation.
C. Transfer Learning
The paper uses two-tier transfer learning to initialize FCNs for DFU segmentation, first from ImageNet and then from Pascal VOC before training on the DFU dataset.
- ImageNet pretraining initializes the convolutional layers using models trained on millions of images.
- Pascal VOC segmentation pretraining follows ImageNet transfer learning and provides initialization for all network layers.
- The resulting pretrained FCNs are then trained on the DFU dataset to improve weight convergence compared with random initialization.
1) FCN-AlexNet:
FCN-AlexNet adapts AlexNet into a fully convolutional architecture for dense segmentation, using convolutionalized layers and deconvolutional upsampling to produce pixel-wise predictions.
- FCN-AlexNet is a fully convolutional version of AlexNet modified for image segmentation.
- Fully connected layers are replaced with equivalent convolutional layers so positional information can be retained for segmentation.
- Deconvolutional layers upsample extracted high-level features to generate a pixel-wise prediction for each input pixel.
- The model uses 500×500 foot images and ground-truth images in Pascal VOC format for training and testing.
2) FCN-32s, FCN-16s, FCN-8s:
The FCN-32s, FCN-16s, and FCN-8s models adapt VGG-16 with different upsampling pathways, producing progressively finer-grained segmentation predictions.
- FCN-32s, FCN-16s, and FCN-8s: FCN-32s, FCN-16s, and FCN-8s are VGG-16-based FCN models with customized upsampling layers.
- FCN-32s: FCN-32s performs end-to-end deconvolution with a 32-pixel stride after convolutionalizing the fully connected layers.
- FCN-16s: FCN-16s combines pool4 and convolutional-layer-7 upsampling to predict on 16×16 pixel blocks.
- FCN-8s: FCN-8s combines pool3, pool4, and convolutional-layer-7 upsampling to predict on 8×8 pixel blocks.
III. EXPERIMENT AND RESULT
The experiments evaluate FCN architectures for complete-area, ulcer, and surrounding-skin segmentation using the reported metrics. FCN-16s leads ulcer and surrounding-skin Dice, while FCN-32s leads complete-area performance.
- The evaluation covers complete area, DFU region, and surrounding skin using Dice, Sensitivity, Specificity, and Matthews Correlation Coefficient.MP denotes model predictions and GT denotes ground-truth labels.
- 0.794 Dice for the ulcer region and 0.851 Dice for surrounding skin were achieved by FCN-16s.FCN-16s was the best performer for these two regions.
- Figure 4 visualizes four examples from different FCN models, with green denoting DFU and red denoting surrounding skin.
- Complete-region segmentation performed better than ulcer and surrounding-skin segmentation in Dice and MCC.
- Healthy-foot testing produced specificity of 1.0 because neither DFU nor surrounding skin was detected.
A. Inaccurate segmentation cases in FCN-AlexNet, FCN-32s, FCN-16s, FCN-8s
Some FCN models produced inaccurate segmentations, particularly for small ulcers, distinct surrounding skin, and irregular boundaries. Smaller pixel strides helped FCN-16s and FCN-8s produce more irregular contours.
- FCN-AlexNet and FCN-32s sometimes missed small DFUs or distinct surrounding skin, or detected only small portions.
- Irregular DFU and surrounding-skin boundaries challenged models that tended to draw more regular contours.
- FCN-16s and FCN-8s produced more irregular contours using smaller pixel strides.
- Some test images contained overlapping DFU and surrounding-skin regions because distinct tissues could resemble the other category.
IV. CONCLUSION
The paper develops FCN-based deep learning approaches for automatic DFU and surrounding-skin segmentation, with potential applications to other lesions and future clinical tools.
- Deep learning FCNs automatically detect and segment DFU and surrounding skin areas with high accuracy.
- The approach may extend to segmenting other skin lesions and classifying wound pathologies.
- The work lays foundations for future multi-class pathology segmentation, automatic annotation, mobile applications, and computer-vision-assisted telemedicine.
- Dice Similarity Coefficient distributions reveal occasional low-scoring cases among the trained models.