Source-linked AI summary
DFUNet: Convolutional Neural Networks for Diabetic Foot Ulcer Classification
Manu Goyal, Neil D. Reeves, Adrian K. Davison, Satyan Rajbhandari, Jennifer Spragg, Moi Hoon Yap
TL;DR
DFU can lead to amputation and imposes costly, lengthy diagnosis and care, motivating more convenient detection. The paper trains conventional classifiers and CNNs, proposes DFUNet, and reports an AUC of 0.962 with 10-fold cross-validation. The authors conclude that automated DFU detection could support remote evaluation and potentially transform diabetic-foot care.
Problem
DFU diagnosis and management rely on costly, lengthy clinical vigilance, while distinguishing ulcerated from healthy skin is challenging because of similarity and variation.
Method
The paper collects expert-labelled DFU and healthy-skin patches and compares conventional machine-learning features with CNNs, including the proposed DFUNet.
Results
DFUNet achieved Sensitivity 0.934, F-measure 0.939, and AUC 0.962, with CNNs generally outperforming conventional machine-learning features.
Takeaways & Limitations
DFUNet enables automated detection of DFU in foot images and has potential for remote DFU evaluation and changes in diabetic-foot care.
Abstract
from arXiv · showhide
Globally, in 2016, one out of eleven adults suffered from Diabetes Mellitus. Diabetic Foot Ulcers (DFU) are a major complication of this disease, which if not managed properly can lead to amputation. Current clinical approaches to DFU treatment rely on patient and clinician vigilance, which has significant limitations such as the high cost involved in the diagnosis, treatment and lengthy care of the DFU. We collected an extensive dataset of foot images, which contain DFU from different patients. In this paper, we have proposed the use of traditional computer vision features for detecting foot ulcers among diabetic patients, which represent a cost-effective, remote and convenient healthcare solution. Furthermore, we used Convolutional Neural Networks (CNNs) for the first time in DFU classification. We have proposed a novel convolutional neural network architecture, DFUNet, with better feature extraction to identify the feature differences between healthy skin and the DFU. Using 10-fold cross-validation, DFUNet achieved an AUC score of 0.962. This outperformed both the machine learning and deep learning classifiers we have tested. Here we present the development of a novel and highly sensitive DFUNet for objectively detecting the presence of DFUs. This novel approach has the potential to deliver a paradigm shift in diabetic foot care.
I. INTRODUCTION
Diabetes and DFUs create substantial health and financial burdens, while current DFU evaluation requires lengthy, specialist-dependent clinical assessment. The paper addresses classification challenges by testing machine-learning and CNN approaches, proposing DFUNet as a faster classifier.
- 15%-25% of diabetic patients may eventually develop DFU, and more than 1 million patients lose part of a leg annually after DFU is not appropriately recognized and treated.
- Current DFU evaluation includes medical-history assessment, specialist examination, and sometimes CT, MRI, or X-Ray testing.
- DFU classification is challenged by costly expert image labelling, high similarity between healthy and ulcerated skin, intraclass variation, lighting, and patient ethnicity.
- The paper tests Conventional Machine Learning methods and CNNs for classifying ulcer versus non-ulcer skin.
- DFUNet is proposed as a novel fast CNN architecture that outperformed GoogLeNet and AlexNet in accuracy and sensitivity.
A. Telemedicine Systems for DFU
Telemedicine can extend DFU assessment beyond conventional care, but existing systems remain device-intensive, specialist-dependent, or insufficiently validated. This motivates automated image-based classification using CNNs and engineered features.
- Existing remote DFU systems use video conferencing, 3D wound imaging, digital photography, or optical scanning but still require specialists for assessment.
- An earlier spectral, infrared, and 3D system required several expensive devices and specialist training.
- A cascaded two-stage SVM system reported promising results but lacked validation on a large dataset and used an impractical contact-based image capture box.
- Hand-crafted texture and color features are affected by lighting and patient skin color, motivating CNN-based representations for DFU classification.
- The paper identifies fully automatic CNN-based DFU-versus-normal-skin classification and introduces DFUNet as a contribution.
III. METHODOLOGY
The methodology builds a labelled DFU image-patch dataset from expert-defined regions, applies augmentation and normalization, and compares conventional features with CNN architectures including DFUNet.
- The experiments compare conventional machine-learning descriptors with LeNet, AlexNet, GoogLeNet, and the proposed DFUNet architecture.
- The study uses DFU and healthy-foot images from patients, with an extensive database of 292 ulcerated-foot images and 105 healthy-foot images.
- The dataset contains 397 foot images and 1,679 labelled skin patches, divided into 1,423 training, 84 validation, and 172 testing patches.
- B. Expert Labelling of Images: Medical experts delineated Regions Of Interest around ulcers and extracted normal and abnormal patches from those regions for ground-truth classification.
- Data augmentation combines rotation, flipping, contrast enhancement, color-space variation, and random scaling to increase training data.
- Zero-centre preprocessing and pixel normalization are applied to the obtained patches.
E. Conventional Machine Learning
The paper compares manually designed color and texture descriptors with CNN-based approaches for distinguishing healthy foot skin from diabetic ulcerations. LeNet activations highlight discriminative features even with a relatively small training set.
- Feature descriptors: Color and texture descriptors were selected as visual cues for distinguishing DFU from healthy skin.SMO was used as the SVM-based classifier for this two-class task.
- Feature descriptors: LBP features capture sudden texture changes in abnormal foot regions for DFU detection.
- Feature descriptors: HOG converts pixel representations into gradients characterized by magnitude and direction.These gradients describe intensity changes at abnormal image locations.
- Convolutional neural networks: LeNet represented diabetic ulcer structures better than traditional features on a relatively small training set.The reported training set contained 1423 patches, with 84 validation patches.
- Convolutional neural networks: The study also evaluated AlexNet and GoogLeNet alongside LeNet for healthy-skin and DFU classification.AlexNet used adjustments for the two-class task and a pre-trained model, while GoogLeNet used inception modules for multilevel feature extraction.
G. Proposed Method - Diabetic Foot Ulcer Network
DFUNet is a CNN architecture designed to improve discriminative feature extraction for diabetic foot ulcer classification. It combines initial traditional convolutions with parallel convolutional layers and a classification output.
- Architecture: DFUNet combines network depth with parallel convolutional layers to extract important DFU features.The architecture begins with traditional convolutional layers and follows them with parallel layers for multiple-feature extraction.
- Feature motivation: DFUNet targets visual differences between smooth healthy skin and ulcers marked by edges and rapid intensity or color changes.
- Architecture: The network has initialization layers inspired by GoogLeNet, parallel convolution layers, and fully connected layers with a softmax-based output.
- Training: DFUNet training used 40 epochs, batch size 8, Adam optimization, and a learning rate of 0.001 with step-down scheduling.The learning rate was reduced by a 33% step with gamma set to 0.1.
2) Parallel Convolutions:
DFUNet uses parallel convolutional layers to extract multiple feature levels from the same input, improving discrimination between healthy skin and DFU.
- Parallel convolution design: Parallel convolution layers concatenate multiple filter outputs from the same input for multilevel feature extraction.This design also covers more spatially spread feature clusters.
- Activation behavior: Each convolution adds discriminative power, with lower activations in healthy skin and higher activations in ulcer images.The authors attribute higher ulcer activations to skin abnormalities.
- Activation and normalization: ReLU layers threshold activations at zero before local response normalization is applied after convolution concatenations.The normalization is described as helpful for avoiding CNN over-fitting.
- Activation and normalization: The normalization output uses source kernel responses and hyperparameters including N, n, α, β, and k.N denotes the total number of kernels, while n denotes the normalization-neighborhood size.
- Dimensionality reduction: Max pooling follows the first and third parallel convolutions to reduce dimensionality.
3) Fully Connected Layers and Output Classifier:
DFUNet ends with average pooling, fully connected layers, and a softmax classifier that outputs probabilities for healthy skin and DFU.
- Output classifier: The output section produces two classes: healthy skin and DFU.Its softmax output represents class probabilities relative to ground-truth labels during training and validation.
- Fully connected layers: Average pooling is followed by fully connected layers with outputs of 100 and 2.The first layer is reduced from 1000 outputs and the last layer is adjusted to two classes for this task.
- Fully connected layers: Fine-tuning the fully connected dimensions is intended to accelerate DFUNet training and testing.
- Output classifier: Softmax converts class scores into values between zero and one that sum to one.The associated loss is defined so good training predictions correspond to a small loss.
IV. RESULTS AND DISCUSSION
DFUNet was evaluated against multiple variants, conventional machine-learning methods, and CNNs using cross-validation and several classification metrics. The proposed model achieved strong sensitivity, F-measure, and AUC while reducing training and testing time relative to GoogLeNet.
- The DFU dataset used an 85% training, 5% validation, and 10% testing split with 10-fold cross-validation.Training and validation used approximately 1,423 and 84 patches, respectively, from 397 original foot images.
- DFUNet variant 5 performed best across evaluation metrics except Precision, where variant 1 was best.Variant 5 used larger filter sizes in its last two parallel convolutional layers.
- CNNs generally outperformed conventional machine-learning features, with GoogLeNet and DFUNet among the best performers across metrics.The comparison included three conventional machine-learning models and CNN architectures including LeNet, AlexNet, GoogLeNet, and DFUNet.
- DFUNet trained in 3 minutes 32 seconds and tested in 49 seconds, compared with GoogLeNet at 16 minutes 27 seconds and 72 seconds.The authors attribute the reduced processing time to the bespoke 14-layer architecture and parallel convolution layers.
- DFUNet achieved Sensitivity 0.934, F-measure 0.939, and AUC 0.962.These were the highest reported performance measures for the proposed DFUNet.
- The experiments found no evidence that lighting conditions or patient ethnicity-related skin tone influenced DFU classification substantially.The authors report very few misclassified testing instances under very high red skin tone.
A. Accurate and Inaccurate Cases of Classification by Proposed DFUNet
DFUNet correctly classified most testing instances but showed specific errors involving subtle or small ulcers and certain normal-skin appearances.
- DFUNet correctly classified most testing instances across abnormal and normal classes.
- Pre-ulcer skin and very small DFUs were often classified as normal.These errors were reported among misclassified abnormal-class examples.
- Normal-skin patches containing toes, highly wrinkled skin, or very high red tone were sometimes classified incorrectly.
V. PERFORMANCE EVALUATION ON HETEROGENEOUS TEST CASE
The study evaluated DFUNet on a heterogeneous dataset captured with an iPad camera after the original DFU dataset had been captured using the same DSLR camera.
- A heterogeneous standardized DFU dataset was collected using an iPad camera to test DFUNet beyond the original single-camera dataset.The authors note that heterogeneous image capture is preferable for computer-vision datasets, but medical ethics constrained use of different cameras initially.
VI. PERFORMANCE EVALUATION ON FACIAL SKIN DATASET
DFUNet was also tested on a three-class facial skin-patch dataset and outperformed GoogLeNet under the same 10-fold cross-validation setting. The authors position the framework as lightweight and potentially extensible beyond two-class DFU classification.
- The facial skin experiment classified normal, spot, and wrinkle patches using GoogLeNet and DFUNet.The dataset contained 110 patches per class.
- DFUNet outperformed GoogLeNet on every evaluation metric in the three-class facial skin experiment.Both architectures used the same experimental settings and 10-fold cross-validation data.
- The authors associate DFUNet’s result with larger filter sizes in later parallel convolution layers that extract multiple features.
- DFUNet is currently fine-tuned for two classes, ulcer and normal skin, with future work targeting additional classes and automated annotation, segmentation, and mobile tools.