[Paper Review] DFUNet: Convolutional Neural Networks for Diabetic Foot Ulcer Classification

Manu S. Goyal, Neil D. Reeves+4 more|arXiv (Cornell University)|Nov 28, 2017

Diabetic Foot Ulcer Assessment and Management4 citations

TL;DR

This paper proposes DFUNet, a lightweight convolutional neural network designed for automated diabetic foot ulcer (DFU) classification in foot images. By leveraging custom architecture with larger filter sizes for enhanced feature extraction, DFUNet achieves an AUC of 0.961 using 10-fold cross-validation, outperforming traditional machine learning and existing CNN models like GoogLeNet and AlexNet in sensitivity and accuracy.

ABSTRACT

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.

Motivation & Objective

• To develop an automated, cost-effective, and remote solution for early detection of diabetic foot ulcers (DFU) using computer vision.
• To address the limitations of current clinical approaches, including high cost, reliance on expert clinicians, and delayed diagnosis.
• To design a novel deep learning architecture, DFUNet, optimized for distinguishing DFU from healthy skin in diverse imaging conditions.
• To evaluate DFUNet’s robustness on heterogeneous datasets, including mobile-captured images and facial skin patches.
• To establish a foundation for future applications in automated ulcer segmentation, multi-class pathology classification, and mobile health tools.

Proposed method

• Proposed a new CNN architecture, DFUNet, with deeper and wider parallel convolutional layers using larger filter sizes to enhance feature extraction for DFU vs. healthy skin classification.
• Trained and evaluated DFUNet using a large, expert-annotated dataset of foot images collected under standardized conditions with a DSLR camera.
• Employed 10-fold cross-validation to ensure robust performance estimation and reduce overfitting.
• Tested generalization on a heterogeneous dataset captured with an iPad camera via the FootSnap application to simulate real-world conditions.
• Extended evaluation to a 3-class facial skin patch dataset (normal, spots, wrinkles) to assess robustness beyond DFU classification.
• Compared DFUNet against established models (GoogLeNet, AlexNet) and traditional machine learning methods using standard metrics including AUC, sensitivity, specificity, precision, F-measure, and MCC.

Experimental results

Research questions

• RQ1 Can a custom-designed CNN architecture outperform existing deep learning models in classifying diabetic foot ulcers from healthy foot skin in clinical images?
• RQ2 How does DFUNet perform on heterogeneous image datasets captured with different devices, such as smartphones, compared to standardized DSLR images?
• RQ3 To what extent can DFUNet generalize to other skin lesion classification tasks, such as distinguishing normal skin, spots, and wrinkles?
• RQ4 What is the impact of architectural design choices—particularly larger filter sizes and reduced depth—on classification performance and inference speed?
• RQ5 Can DFUNet achieve high sensitivity and specificity in a low-resource setting, enabling remote and scalable DFU screening?

Key findings

• DFUNet achieved an AUC score of 0.961 on the primary dataset using 10-fold cross-validation, significantly outperforming both traditional machine learning and deep learning baselines.
• On a heterogeneous test set captured with an iPad camera, DFUNet achieved a sensitivity of 0.929, specificity of 0.908, precision of 0.942, F-measure of 0.931, and AUC of 0.950.
• In a 3-class facial skin classification task, DFUNet outperformed GoogLeNet with a sensitivity of 0.867 vs. 0.783, specificity of 0.930 vs. 0.882, and F-measure of 0.867 vs. 0.784.
• The model demonstrated high robustness to variations in image capture conditions, including lighting and device type, indicating strong generalization potential.
• Despite using fewer layers and a lightweight design, DFUNet achieved higher sensitivity and specificity than deeper models like GoogLeNet and AlexNet.
• Misclassifications primarily occurred in small ulcers and regions with high redness or wrinkled skin, indicating areas for future architectural refinement.