PaperlayerView exact revision on arXiv

Source-linked AI summary

A Machine Learning Model for Early Detection of Diabetic Foot using Thermogram Images

Amith Khandakar, Muhammad E. H. Chowdhury, Mamun Bin Ibne Reaz, Sawal Hamid Md Ali, Md Anwarul Hasan, Serkan Kiranyaz, Tawsifur Rahman, Rashad Alfkey, Ahmad Ashrif A. Bakar, Rayaz A. Malik

arXiv:2106.14207v1eess.IVcs.CVcs.LG

TL;DR

DFU can be difficult or delayed to heal, and bilateral thermal changes can complicate asymmetry-based assessment. The paper compares feature-based machine learning with 2D CNN thermogram classification and reports evaluations of foot-ulcer detection and binary single-foot classification.

  • Problem

    DFU can be difficult or delayed to heal and raises infection and amputation risk; bilateral foot changes can undermine asymmetry-based comparison.

    Source: 1.Introduction · p. 2

  • Method

    The paper compares 2D deep-learning thermogram classification with feature selection, optimization, and traditional classifiers.

    Source: 1.Introduction · p. 4

  • Results

    Results address foot-ulcer detection and binary classification of single-foot thermograms using 2D CNNs.

    Source: §3 · p. 13

  • Takeaways & Limitations

    The authors investigate relevant features intended to improve diabetic-foot detection when used as inputs to traditional classifiers.

    Source: 1.Introduction · p. 4

    Bilateral thermal changes may prevent one foot serving as an asymmetry reference, and cited prior deep learning was not considered suitable for clinical decisions or discussed for smartphone use.

    Source: 1.Introduction · pp. 2–3

Abstract

from arXiv · show

Diabetes foot ulceration (DFU) and amputation are a cause of significant morbidity. The prevention of DFU may be achieved by the identification of patients at risk of DFU and the institution of preventative measures through education and offloading. Several studies have reported that thermogram images may help to detect an increase in plantar temperature prior to DFU. However, the distribution of plantar temperature may be heterogeneous, making it difficult to quantify and utilize to predict outcomes. We have compared a machine learning-based scoring technique with feature selection and optimization techniques and learning classifiers to several state-of-the-art Convolutional Neural Networks (CNNs) on foot thermogram images and propose a robust solution to identify the diabetic foot. A comparatively shallow CNN model, MobilenetV2 achieved an F1 score of ~95% for a two-feet thermogram image-based classification and the AdaBoost Classifier used 10 features and achieved an F1 score of 97 %. A comparison of the inference time for the best-performing networks confirmed that the proposed algorithm can be deployed as a smartphone application to allow the user to monitor the progression of the DFU in a home setting.

1.Introduction

Early detection is challenging because diabetic-foot thermal patterns vary and bilateral changes can undermine asymmetry-based comparison. The paper compares thermogram classification using feature-based machine learning and 2D deep learning.

  • DFU can be difficult or delayed to heal, raises infection and amputation risk, and recurs in approximately 40% of patients after one year and 60% after three years.Source: 1.Introduction · p. 2
  • Prior temperature-monitoring studies reported predicting DFU in 97% of patients and used a 2.22°C difference over at least two consecutive days as a therapy threshold.Source: 1.Introduction · p. 2
  • Thermography rapidly and non-invasively quantifies diabetic-foot thermal changes, but bilateral changes can prevent the opposite foot from serving as an asymmetry reference.The limitation applies when both feet show thermal changes without a butterfly pattern.Source: 1.Introduction · p. 2
  • Prior studies extracted plantar hot-region features; reported approaches included active contours, Discrete Wavelet Transformation with higher-order spectra, and image decomposition via DD-DT-CWT.Source: 1.Introduction · p. 3
  • Earlier methods often used small private datasets, and their generalizability and real-time inference were not established [2].The passage also notes that their performance was not comparable to machine-learning-based techniques.Source: 1.Introduction · p. 3
  • The authors use an available dataset to classify control and diabetic groups and compare their automatic thermogram classifier with 2D deep learning.Source: 1.Introduction · p. 4
  • The study evaluates state-of-the-art 2D CNN models and image-enhancement techniques for diabetic-foot detection [40-43].Source: 1.Introduction · p. 4
  • The authors investigate relevant features for traditional classifiers and compare feature-selection, optimization, and classification techniques using lightweight classifiers.Source: 1.Introduction · p. 4

2. Methodology

The methodology uses thermogram images to distinguish diabetic from control feet through engineered temperature-map features and 2D CNN classification. It covers preprocessing, feature optimization, classifier evaluation, and measures of predictive performance and inference time.

2.3.Thermogram Image Classification by 2D CNNs

  • The image-classification arm evaluated 2D CNNs on thermogram images, with image-enhancement techniques used to improve performance and five-fold cross-validation applied.The supplied passage introduces the cross-validation procedure but does not provide further CNN setup details.Source: §2 · p. 4, 2.3.Thermogram Image Classification by… · p. 11

2. Methodology

  • Thermograms from 167 foot pairs from 122 people with diabetes and 45 controls were segmented into four angiosomes and used to classify diabetic versus control feet.The dataset included age, gender, height, and weight; the source dataset was made public by Hernandez-Contreras et al..Source: §2 · pp. 4–5

2.2.Classification using thermogram features

  • Feature-based classification used five-fold cross-validation, an 80/20 training-test split, a validation subset, and SMOTE augmentation of training data.Twenty percent of training data was reserved for validation.Source: 2.2.Classification using thermogram… · p. 8
  • Features with more than 95% correlation were removed from the 39-feature set to reduce redundancy and overfitting.The reduced set contained 28 features.Source: §2.1 · p. 7, 2.2.Classification using thermogram… · p. 9
  • The feature-based search combined three feature-selection techniques with 10 classifiers in 840 investigations using the optimized 28-feature set.The three ranking approaches were XGBoost, Random Forest, and Extra Tree.Source: 2.2.Classification using thermogram… · p. 10

2.3.Thermogram Image Classification by 2D CNNs

  • Six ImageNet-pretrained CNNs were evaluated using five-fold assessment, with image rotation and translation augmenting the small, imbalanced training data.Performance was reported cumulatively across folds.Source: 2.3.Thermogram Image Classification by… · p. 11
  • Adaptive histogram equalization and gamma correction were applied to thermograms to enhance local contrast and pixel values.AHE equalizes histograms over image patches, adapting local contrast and edges to pixel-intensity distributions.Source: 2.3.Thermogram Image Classification by… · p. 12

2.4 Performance metrics

  • Six metrics—Sensitivity, Specificity, Precision, Accuracy, F1-Score, and AUC—were used; overall accuracy and weighted performance metrics were reported, with 95% confidence intervals for four measures.Confidence intervals were reported for Sensitivity, Specificity, Precision, and F1 Score.Source: §2.4 · p. 12
  • Inference time was calculated for the best-performing 2D CNNs and 1D classifiers to compare computational complexity and identify models suitable for smartphone deployment.Source: §2.4 · pp. 12–13

3. Results and Discussion

The experiments compare deep CNNs across thermogram configurations and image enhancement methods with classifiers trained on optimized thermogram features. AdaBoost with Random Forest feature ranking performs best among the feature-based methods and outperforms 2D image-based deep learning, while MobileNetV2 is the strongest reported CNN for dual-foot thermograms.

3. Results and Discussion

  • The results compare transfer-learned CNNs across enhancement methods and single- versus dual-foot inputs, alongside classifiers evaluated on optimized thermogram features.The feature-based classifiers were evaluated with different feature-selection techniques and feature combinations.Source: §3 · p. 13

3.1 Detection results by deep CNN models

  • 94.01% overall sensitivity: DenseNet201 performed best among the six CNNs for single-foot thermograms, with class-wise sensitivities of 95.9% for DM and 88.89% for CG.Original single-foot thermograms performed better than AHE- and Gamma-enhanced versions.Source: Table IV · p. 14, §3.1 · p. 14
  • 95.81% overall sensitivity: MobileNetV2 was the best-performing network for dual-foot thermograms, with class-wise sensitivities of 96.72% for DM and 93.33% for CG.The paper reports these results for the Gamma-enhanced dual-foot input.Source: Table V · p. 14, §3.1 · p. 15
  • Gamma enhancement improved classification over original thermograms for dual-foot investigation, while combined-foot images provided more distinguishable features.The reported comparison is specific to the dual-foot CNN setting.Source: §3.1 · p. 15, Figure 5 · p. 15

3.2 Feature-based detection results

  • 96.71% sensitivity was achieved using feature selection with the top 10 features, and AdaBoost with Random Forest ranking outperformed 2D image-based deep learning.The top 10 features included age, temperature-variation measures, and thermogram-derived indicators.Source: §3.2 · pp. 16, 18, 19, Table VII · p. 16
  • Only MobileNetV2 among the CNNs and AdaBoost among the traditional classifiers were identified as deployable on smart portable devices.Figure 6 compares F1-score with inference time for top-performing techniques.Source: Figure 6 · p. 16, §3.2 · p. 17

V. CONCLUSION

The paper proposes thermogram-based early detection of the diabetic foot, motivated by the morbidity associated with diabetic foot ulceration and amputation. Its classical machine-learning framework uses optimized features from foot thermograms, and the authors report that it outperforms 2D image-based deep learning; clinical-trial validation is noted as a next step.

V. CONCLUSION

  • Diabetic foot ulceration has a major impact on morbidity and mortality, and early detection may help limit ulcer progression and eventual amputation.The authors identify potential utility for health professionals, caregivers, and patients tracking the disease.Source: §V · p. 19
  • The proposed diagnostic system based on optimized thermogram features outperformed 2D image-based deep learning.The framework targets early diabetic-foot detection using infrared-camera thermograms captured with a smartphone.Source: §V · pp. 19–20
  • The authors state that the smartphone-based application should be validated in a clinical trial.They also describe the proposed network as suitable for smartphone-based deployment.Source: §V · p. 20
Loading 2106.14207v1…