Source-linked AI summary
Recognition of Ischaemia and Infection in Diabetic Foot Ulcers: Dataset and Techniques
Manu Goyal, Neil Reeves, Satyan Rajbhandari, Naseer Ahmad, Chuan Wang, Moi Hoon Yap
TL;DR
The paper addresses limited computerized recognition of diabetic-foot-ulcer infection and ischaemia, which inform healing and amputation-risk assessment. It introduces a labelled dataset and combines superpixel colour descriptors, natural region-focused augmentation, and ensemble CNNs; CNNs outperform handcrafted methods, especially for ischaemia.
Problem
Computerized DFU research has mainly focused on visual appearance and tissue classification, while infection and ischaemia recognition remains clinically important for assessing healing and amputation risk.
Method
The paper introduces a DFU dataset with ischaemia and infection labels and evaluates superpixel colour descriptors, natural region-focused augmentation, and an Ensemble CNN combining Inception-V3, ResNet50, and InceptionResNetV2 features.
Results
CNNs outperformed traditional machine-learning methods on both tasks, with higher average accuracy for ischaemia than infection and the Ensemble CNN achieving the highest scores across measures.
Takeaways & Limitations
The study supports image-based machine learning as a proof of concept for DFU ischaemia and infection classification, with stronger performance for ischaemia than infection.
Takeaways & Limitations
Performance may be constrained by image-based ground truth and dataset scale; future work should use vascular and blood-test assessments and larger datasets.
Abstract
from arXiv · showhide
Recognition and analysis of Diabetic Foot Ulcers (DFU) using computerized methods is an emerging research area with the evolution of image-based machine learning algorithms. Existing research using visual computerized methods mainly focuses on recognition, detection, and segmentation of the visual appearance of the DFU as well as tissue classification. According to DFU medical classification systems, the presence of infection (bacteria in the wound) and ischaemia (inadequate blood supply) has important clinical implications for DFU assessment, which are used to predict the risk of amputation. In this work, we propose a new dataset and computer vision techniques to identify the presence of infection and ischaemia in DFU. This is the first time a DFU dataset with ground truth labels of ischaemia and infection cases is introduced for research purposes. For the handcrafted machine learning approach, we propose a new feature descriptor, namely the Superpixel Color Descriptor. Then we use the Ensemble Convolutional Neural Network (CNN) model for more effective recognition of ischaemia and infection. We propose to use a natural data-augmentation method, which identifies the region of interest on foot images and focuses on finding the salient features existing in this area. Finally, we evaluate the performance of our proposed techniques on binary classification, i.e. ischaemia versus non-ischaemia and infection versus non-infection. Overall, our method performed better in the classification of ischaemia than infection. We found that our proposed Ensemble CNN deep learning algorithms performed better for both classification tasks as compared to handcrafted machine learning algorithms, with 90% accuracy in ischaemia classification and 73% in infection classification.
1. Introduction
DFU infection and ischaemia are clinically important conditions linked to healing and amputation risk, yet image-based recognition remains difficult. The paper motivates machine-learning methods to identify these conditions from foot images despite dataset and clinical-assessment challenges.
- DFUs can lead to foot or limb amputation, making recognition of infection and ischaemia important for assessing healing and amputation risk.
- Ischaemia is inadequate blood supply that can impair DFU healing and may appear visually as poor reperfusion or gangrenous toes.
- Bacterial infection is defined by bacterial soft-tissue or bone infection, with redness and purulence providing possible visual indications.
- Clinicians show low validity and reliability when remotely assessing DFUs from foot images, underscoring the difficulty of image-based assessment.
- Machine-learning assessment is challenged by labelling burdens, class similarity and variation, non-standardized imaging, and missing patient metadata.
- The paper aims to replicate experts’ visual predictions of ischaemia and infection using cost-effective machine learning to support future remote DFU monitoring.
2. DFU Dataset and Expert Labelling
The paper introduces a 1459-image DFU dataset with expert binary labels for ischaemia and infection. The images lack clinical metadata, so the study relies on visual expert annotation and image-based learning.
- 1459 DFU foot images were collected over five years at Lancashire Teaching Hospitals with ethical approval and informed consent.
- The dataset contains images captured using Kodak DX4530, Nikon D3300, and Nikon COOLPIX P100 cameras.
- The dataset’s NHS-approved records contained no clinical metadata or medical classification for ischaemia and infection.
- The dataset was introduced as the first DFU image collection with ground-truth labels for ischaemia and infection.
- Two diabetic-foot consultant physicians visually labelled ischaemia and infection, with a senior physician resolving disagreements.
3. Methodology
The methodology combines ROI-focused Natural Data-augmentation, handcrafted Superpixel Color Descriptors, and fine-tuned or ensemble CNNs for binary recognition of DFU ischaemia and infection.
- 3.1. Natural Data-Augmentation Technique based on Deep Learning Algorithm: Natural Data-augmentation uses Faster R-CNN with InceptionResNetV2 to localize DFU regions before applying adjustable magnifications and augmentations.Three magnifications are used in the experiment, with magnification distance and number of factors adjustable.
- 3.1. Natural Data-Augmentation Technique based on Deep Learning Algorithm: The dataset contains imbalanced ischaemia classes, with 1249 ischaemia and 210 non-ischaemia cases, while infection classes contain 628 infection and 831 non-infection cases.Binary classification is used because the multi-class distribution is especially imbalanced for ischaemia and no infection cases.
- 3.2. Handcrafted Superpixel Color Descriptors: The Superpixel Color Descriptor extracts red and black color regions from DFU patches after SLIC over-segmentation and per-superpixel mean RGB computation.The selected configuration uses k=200 superpixels and threshold values for red and black intensities, producing a 10-dimensional feature vector.
- 3.2. Handcrafted Superpixel Color Descriptors: Random Forest and Multilayer Perceptron achieved the highest accuracy among the tested traditional classifiers using the handcrafted color features.The approach was developed after texture and conventional color descriptors produced very poor results for the binary tasks.
- 3.3. Deep Learning Approaches: The Ensemble CNN combines bottleneck features from Inception-V3, ResNet50, and InceptionResNetV2, then uses an SVM classifier to produce predictions.The CNN models are fine-tuned through transfer learning, with early layers initially frozen for common features.
4. Results and Discussion
The experiments compare traditional machine learning and CNN approaches for binary ischaemia and infection classification, using localized DFU patches and cross-validation. CNNs, especially the Ensemble CNN, performed better overall, with ischaemia easier to classify than infection; image variability and subtle visual indicators remained important limitations.
- Experimental setup: The experiments used 70% training, 10% validation, and 20% testing splits, 5-fold cross-validation, and natural data augmentation for both classification tasks.The ischaemia dataset produced approximately 11,564 training, 1,652 validation, and 3,304 testing patches.
- CNN approach: The Ensemble CNN combines bottleneck features from Inception-V3, ResNet50, and InceptionResNetV2, then uses an SVM classifier for predictions.The approach extracts features from multiple CNN models rather than relying on a single CNN architecture.
- Overall performance: 83.3% average accuracy was achieved for ischaemia classification, compared with 65.8% for infection classification across the evaluated models.Average MCC and AUC were 67.1% and 83.2% for ischaemia, versus 32.3% and 65.8% for infection.
- Model comparison: 86.5% CNN accuracy exceeded 79% for traditional machine learning in ischaemia classification, while 68.4% exceeded 62.1% in infection classification.The Ensemble CNN achieved the highest score in all reported performance measures for both tasks.
- Performance measures: The Ensemble CNN achieved 92.9% specificity for ischaemia, narrowly exceeding ResNet50 at 92.7%.For infection, it achieved 74.4% specificity and 70.9% sensitivity, exceeding the other CNN classifiers on the reported measures.
- Error analysis: Misclassification was associated with intra-class dissimilarity, inter-class similarity, shadows, skin tone, lighting, and visual indicators that were too subtle to recognize.The authors also suggest that more sensitive vascular or blood-test ground truth may be needed, particularly for infection.
5. Conclusion
The paper reports binary classification of ischaemia and infection in DFUs using traditional machine learning and CNNs, while identifying challenges that constrain performance. It concludes that better-balanced data, improved image capture, and clinically informed ground truths could improve future results.
- The study trained traditional machine learning classifiers and CNNs to distinguish ischaemia from non-ischaemia and infection from non-infection in DFUs.
- Visual similarity between classes and subtle indicators of infection and ischaemia make classification from nonstandard foot images challenging.
- More balanced datasets, improved data capture, and clinical tests providing ground truths could improve algorithm performance and sensitivity or specificity.