Hybrid Vae‑XGBoost Framework for Efficient Classification of
Diabetic Foot Ulcer Images
N. Nagarani, Gokul Priyan G. V., Sivanesan R. and Sukha Dev A.
Velammal College of Engineering and Technology, Madurai, Tamil Nadu, India
Keywords: Diabetic Foot Ulcer, Variational Autoencoder, XGBoost, Feature Extraction, Classification Accuracy.
Abstract: Diabetic Foot Ulcer (DFU) classification is critical for early classify and planning for manage, with a view to
minimizing complications. In this article, a new hybrid model is developed, with Variational Autoencoder
(VAE) for feature extraction and XGBoost for classify and with a view to improving accuracy and efficiency
in classify of DFU images. VAE learns a low-dimensional and discriminative feature representation of ulcer
images, encoding significant structures and textures and dimensionality reduction. Features extracted via VAE
are then fed into an optimized XGBoost classify, with a view to improving decision-making via gradient-
boosted trees. The proposed model is compared with a benchmarked DFU dataset and contrasted with
traditional deep networks, with considerable performance improvement in accuracy, precision, recall, and F1-
score. Experimental observations confirm that combining VAE for unsupervised feature extraction with
XGBoost for classify enormously improves robustness and generalizability. This hybrid model introduces an
efficient and interpretable model for computerized DFU classify, with a view to supporting clinicians in early
and correct classify.
1 INTRODUCTION
Diabetic Foot Ulcer (DFU) is one of the most severe
diabetes mellitus complications, and it occurs in
millions of patients worldwide. Approximately 15%
of the diabetes patients get ulcers in the foot, which,
if untreated, can go on to get infected, result in
gangrene, and lead to the loss of a limb. Having a
presence of DFUs in a high chance of getting a patient
hospitalized and even dead, and thus, early detection
and proper grading become critical for effective
therapy and averts grave complications. Traditional
DFU diagnosis consists of clinical examination,
estimation of wound depth, and radiologic modalities
such as infrared thermography and Doppler
ultrasound. Grading scales such as Wagner
Classification System and Texas University Wound
Classification have been used for estimating severity
of an ulcer, but such techniques are subjective, time
consuming, and have inter-observer variation, and
thus, computerized grading tools become a necessity.
In recent years, methodologies in Machine
Learning (ML) and Deep Learning (DL) have been
powerful tools for DFU image classification, with
high accuracy and efficiency over conventional,
manual methodologies. Methods such as Support
Vector Machines (SVM), Random Forest, and
XGBoost have been adopted for DFU classification,
using hand-designed features such as texture, color,
and shape descriptors. However, such methodologies
have been restricted by the need for feature
engineering, a process that sometimes fails to extract
complex visual structures in DFUs. Convolutional
Neural Networks (CNNs), ResNet, VGG, and
EfficientNet, under deep learning, have been seen to
outdo them through a capability to learn
discriminative features in an unsupervised manner
directly from raw DFU images. In contrast, even with
success, deep networks require a lot of labelled data,
use a lot of computation, and suffer from overfitting,
specifically when dealing with small, unbalanced
medical datasets.
To overcome such challenges, in this work, a
Hybrid VAE-XGBoost model is proposed, leveraging
the capabilities of Variational Autoencoder (VAE) for
unsupervised feature extraction and XGBoost for
efficient classification of DFU images. VAE model is
adopted for discovering a concise, reduced-
dimensional abstraction of DFU images, with reduced
dimensions and retained important structures and
Nagarani, N., V., G. P. G., R., S. and A., S. D.
Hybrid Vae-XGBoost Framework for Efficient Classification of Diabetic Foot Ulcer Images.
DOI: 10.5220/0013895500004919
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages
209-214
ISBN: 978-989-758-777-1
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
209
textures information. High-dimensional
representations extracted through them are then
leveraged for training an optimized XGBoost
classifier, famous for its efficiency in processing
structured information and producing strong
prediction with less overfitting. By fusing deep
feature extraction with a high-performance gradient-
boosted decision tree classifier, proposed model
achieves increased accuracy, generalizability, and less
computational cost compared with standalone deep
learning and traditional ML models.
The primary contributions of this study include:
(1) Proposing a VAE-XGBoost model that achieves
high accuracy and efficiency in the classification of
DFU. (2) Investigating representations within the
latent space to enhance features in DFU while
minimizing reliance on large datasets. (3) Conducting
a comparison with leading deep neural networks,
demonstrating comparable performance in terms of
accuracy, precision, recall, and F1-score. (4)
Presenting a lightweight and transparent model
suitable for real-world medical applications,
facilitating early and computerized diagnosis of DFU
for healthcare professionals.
The rest of this paper is organized as follows: Section
2 reviews related research in DFU classification,
Section 3 details the proposed method, Section 4
presents experimental results and comparisons, and
Section 5 concludes with suggestions for future
research and final thoughts.
2 LITERATURE SURVEY
A. Huong et al present an application of an
automatized technique for optimization in finding an
ideal solution for a problem. PSO was utilized in
overcoming a disadvantage of a conventional
technique, and in improving training of a neural
network for its use in diabetic foot ulcer (DFU)
application. The system forms an ideal platform for
technology adaptability in controlling DFU. It can
even act as an ideal decision-support tool for limb
salvage and healing processes' optimizations.
A. Huong, et al utilizes a two-dimensional image
asits basis and a collection of neural networks for
picture processing. It can label an image in four
categories, infection, ischaemia, both, and none.
X.Wuetal developed a flexible model for creating
an efficient augmentation pool for Diabetic Foot
Ulcers medical images. In addition, we use ensemble
learning for enhancing model performance. Unlike
conventional plurality voting, we present a scheme
with a name "voting with expertise" having a bias
towards prediction with reasonably sound value.
Experimental testing confirms efficacy of proposed
techniques and secured a second rank through
integration of aforementioned two enhancements in
present ongoing challenge-Dfuc2021 Challenge.
3 PROPOSED SYSTEM
The novel Hybrid VAE-XGBoost system is created to
classify DFU images better by combining deep
learning-based Variational Autoencoder (VAE) to
perform feature extraction and XGBoost to conduct
classification. Deep learning algorithms such as
CNNs have been known to require immense
computational power and immense training sets,
while traditional machine learning algorithms have
been dependent upon hand-crafted feature extraction
that does not perform in every context. To address
these issues in this work, the system utilizes VAE to
get meaningful DFU image representations in the
latent space. The encoder in the VAE compresses the
input image xxx to get a latent variable zzz following
a Gaussian distribution:
(1)
where μ and σ represent the learned mean and
variance. The decoder then reconstructs the original
image from zzz, ensuring the preservation of crucial
visual information. The loss function of VAE consists
of two components: Rconstruction loss (Lrec), which
minimizes the difference between input and
reconstructed image, given by:
(2)
and the KullbackLeibler (KL) divergence loss LKL ,
which ensures that the learned distribution remains
close to a standard normal distribution:
(3)
where β is employed to balance between
regularization in the latent space and performance in
reconstruction. This is subsequently followed by a
process of selecting features in the form of Principal
Component Analysis (PCA) or statistics-based
importance to remove redundant information while
retaining only the most discriminatory features.
ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,
COMMUNICATION, AND COMPUTING TECHNOLOGIES
210
The desired attributes are then passed to the
XGBoost classifier where an efficient classification is
done by adopting a gradient-boosted decision tree
process. XGBoost minimizes an objective function
that consists not only of an added loss term but an
added regularization term:
(4)
By combining deep feature learning from VAE
with the structured decision-making capability of
XGBoost, the proposed system achieves higher
classification accuracy, reduced computational cost,
and better generalization compared to standalone
deep learning models. This Hybrid VAE-XGBoost
framework is particularly effective for small medical
datasets, as it leverages unsupervised learning to
extract robust representations and gradient boosting
to make optimal predictions. The model performance
is evaluated using metrics such as Approach, Success
Rate (%) Exactness (%), Sensitivity (%), F1-Measure
(%), ensuring its reliability in real-world clinical
applications.
Figure 1: Proposed System Modules.
The proposed Hybrid VAE-XGBoost framework
is structured into multiple modules, ensuring an
efficient workflow from data acquisition to
classification and evaluation. Each module plays a
vital role in improving the accuracy and
generalization of the model for Diabetic Foot Ulcer
(DFU) image classification. The key modules are as
follows: Figure 1 show the Proposed system Modules
3.1 Data Acquisition and Pre-
Processing
This module involves collecting DFU images from
publicly available datasets or hospital-based
repositories. Since medical images often contain
noise, illumination variations, and artifacts, proper
pre-processing techniques are applied to ensure
uniformity and quality before feature extraction.
Key pre-processing steps include:
• Image Resizing: Standardizing images to a
fixed dimension for consistency.
• Normalization: Scaling pixel intensities to a
standard range to improve model stability.
• Data Augmentation: Applying techniques
such as rotation, flipping, contrast
enhancement, and noise addition to increase
dataset variability and reduce overfitting.
• Segmentation (if required): Extracting the
ulcer region using U-Net or thresholding
techniques to focus on relevant features.
This module ensures that the input data is optimized
for feature extraction and classification.
3.2 Feature Extraction Using
Variational Autoencoder (VAE)
In this module, a Variational Autoencoder (VAE) is
utilized to extract significant features from DFU
images. The encoder maps the image into a
condensed, low-dimensional latent space, capturing
vital information and filtering out unnecessary noise
and redundancy. The decoder subsequently
reconstructs the image from this representation,
ensuring that only the most relevant visual elements
are preserved.
VAE is particularly effective in learning robust
and structured representations that are useful for
classification. Instead of using raw pixel values, the
latent space embeddings generated by the encoder
serve as input for the next stage of the pipeline.
3.3 Feature Selection and
Dimensionality Reduction
Since deep learning models often generate high-
dimensional feature spaces, it is crucial to select the
most informative features to enhance classification
efficiency. This module applies Principal Component
Analysis (PCA) or other statistical techniques to
eliminate redundant or less significant features. By
reducing dimensionality, the model ensures faster
training and better generalization while maintaining
important ulcer characteristics.
3.4 Classification Using XGBoost
The refined feature set is then passed into XGBoost,
an optimized gradient boosting algorithm that builds
multiple decision trees to classify images. XGBoost
is chosen due to its efficiency, scalability, and ability
to handle imbalanced datasets. It constructs trees
Hybrid Vae-XGBoost Framework for Efficient Classification of Diabetic Foot Ulcer Images
211
sequentially, with each tree correcting errors made by
the previous one.
During training, XGBoost optimizes
hyperparameters such as learning rate, tree depth, and
number of estimators to improve classification
performance. The classifier outputs the final ulcer
classification, distinguishing between normal skin,
infected ulcer, and healing ulcer based on extracted
features.
3.5 Performance Evaluation and
Validation
To assess the effectiveness of the proposed system,
various performance metrics are calculated,
including:
• Success Rate (%): Measures the overall
correctness of the model.
• Exactness (%) and Sensitivity (%): Evaluate
the model’s ability to correctly classify
ulcers.
• F1-Measure (%): Balances precision and
recall, especially for imbalanced datasets.
• AUC-ROC Curve: Analyzes the classifier’s
ability to distinguish between ulcer types.
Cross-validation techniques, such as k-fold
validation, are applied to ensure that the model
generalizes well to unseen data. The results are then
compared with traditional CNN-based models to
highlight the advantages of the Hybrid VAE-
XGBoost framework.
4 RESULT & DISCUSSION
For this work, we have accumulated a dataset of 800
images from website which was divided in three
phases. We have divided this dataset in training
dataset of 80 images and testing dataset of 20 images.
We have trained the Multi scale architecture in
training dataset by following transfer learning
technique in which pre-trained weights of proposed
model have been used to initialize training weights.
We have trained the VAE model over training dataset
in 50 epochs with batch size=10 and learning rate= 0.
0001.. Figure 2 show the Input image.
Figure 2: Input Image.
Figure 3: Validation and Testing Curve.
Figure 4: Classification Result.
ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,
COMMUNICATION, AND COMPUTING TECHNOLOGIES
212
Figure 3 and 4 shows the Validation and Testing curve
and Classification result respectively.
The training curve represents the model's progress on
the training dataset over time, whereas the validation
curve illustrates its performance on the validation
dataset. Ideally, the model's efficiency should
increase with each epoch until it stabilizes at a certain
level.
Table 1 Performance Analysis of Precision, F1
Score, Accuracy and Specificity of the proposed
method with Various Models
Table 1: Performance Comparison.
Approac
h
Succes
s Rate
(%)
Exactnes
s (%)
Sensitivit
y (%)
F1-
Measur
e (%)
CNN-
Based
Model
85.3
82.7
83.1
82.9
ResNet-
50
88.9
87.5
86.8
87.1
VGG-16
87.2
85.9
85.5
85.7
XGBoos
t (Raw
Features
)
83.5
81.2
80.9
81.0
Propose
d VAE-
XGBoos
t
92.4
91.1
90.5
90.8
5 DISCUSSIONS
Highest accuracy (92.4%) is produced by the Hybrid
VAE-XGBoost framework compared to CNN-based
approaches (ResNet-50, VGG-16) and raw features
Precision, Recall, and F1-score of the resultant model
are significantly improved due to efficient feature
extraction by VAE that learns to identify
discriminative ulcer patterns while reducing noise.
The traditional CNN methods including ResNet-
50 and VGG-16 have accuracies that are high but
have computationally demanding models and
extensive training sets.
The performance of XGBoost trained over raw
features is relatively inferior because hand-crafted
features perform poorer compared to deep-learned
representations extracted by VAE.
The technique utilizes both strengths of deep
feature extraction (VAE) and machine learning
classification (XGBoost) to gain improved
generalizability and stability.
Impact on Medical Diagnosis: The results indicate
that the Hybrid VAE-XGBoost system can
effectively aid healthcare workers in early DFU
identification to prevent amputation and serious
complications. The system provides:
• Enhanced diagnostic specificity to reduce
misclassification
• Successful pattern recognition with respect
to ulcers.
• Reduced overfitting because VAE learns in
an organized feature space.
• Scalability and interpretability to render it
appropriate to apply in clinical practice.
The superior performance of this model suggests that
it can be used in mobile diagnostic apps or in
telemedicine systems to achieve automated, efficient,
and accurate DFU classification.
6 CONCLUSIONS
The developed Hybrid VAE-XGBoost technique
provides an accurate yet efficient DFU classification
method. Using the Variational Autoencoder (VAE) to
perform deep feature extraction and XGBoost to
obtain reliable classification results, the system
provides improved performance in identifying
diverse types of ulcers. Deep learning-based
representation learning in combination with machine
learning-based classification provides improved
generalizability, overfitting minimization, and
diagnostic performance. Strong preprocessing,
feature selection, and evaluation practices further
confirm the reliability of the system. Experimental
results confirm that the developed method is superior
to conventional CNN-based methods in providing an
efficient, scalable, and interpretable DFU diagnostic
method. Future work can focus on integrating real-
world deployment, multi-modal fusion capabilities,
and explainable AI practices to further support
clinical utility.
The system can support early DFU identification
among professionals to avoid future complications
and improved outcomes in patients.
REFERENCES
C. Y. Rubavathi and J. Diofrin, "Diabetes Foot Ulcer
Diagnosis using Fast Convolution Neural Network,"
Hybrid Vae-XGBoost Framework for Efficient Classification of Diabetic Foot Ulcer Images
213
2023 International Conference on Networking and
Communications (ICNWC), Chennai, India, 2023, pp.
1-5, doi: 10.1109/ICNWC57852.2023.10127332.
E. Santos, F. Santos, J. Dallyson, K. Aires, J. M. R. S.
Tavares and R. Veras, "Diabetic Foot Ulcers
Classification using a fine-tuned CNNs Ensemble,"
2022 IEEE 35th International Symposium on
Computer-Based Medical Systems (CBMS), Shenzen,
China, 2022, pp. 282- 287, doi:10.1109/CBMS55023.
2022.00056.
F. Santos et al., "DFU-VGG, a Novel and Improved VGG-
19 Network for Diabetic Foot Ulcer Classification,"
2022 29th International Conference on Systems,
Signals and Image Processing (IWSSIP), Sofia,
Bulgaria, 2022, pp. 1- 4, doi: 10.1109/IWSSIP55020.
2022.9854392
F. Arnia, K. Saddami, R. Muharar, D. A. Dwi Pratiwi and
Y. Nurdin, "Diabetic Foot Ulcer Detection on Mobile
Platforms through Thermal Imaging and Deep
Learning," 2023 International Conference on Smart-
Green Technology in Electrical and Information
Systems (ICSGTEIS), Badung, Bali, Indonesia, 2023,
pp. 104- 108, doi:10.1109/ICSGTEIS60500.2023.104
24364.
H. Jin and L. Liu, "Quantification of Diabetic Foot Ulcer
Based on Ulcer Segmentation," 2022 IEEE
International Conference on Industrial Technology
(ICIT), Shanghai, China, 2022, pp. 1-7, doi:
10.1109/ICIT48603.2022.10002830.
J. Katual and A. Kaul, "Analysis of Thermal Images with
Parallel Convolutional Deep Neural Network for
Diabetic Foot Detection," 2022 IEEE 3rd Global
Conference for Advancement in Technology (GCAT),
Bangalore, India, 2022, pp. 1-5, doi:
10.1109/GCAT55367.2022.9972064.
Kumar, L. Nelson and S. Singh, "ResNet-50 Transfer
Learning Model for Diabetic Foot Ulcer Detection
Using Thermal Images," 2023 2nd International
Conference on Futuristic Technologies (INCOFT),
Belagavi, Karnataka, India, 2023, pp. 1-5, doi:
10.1109/INCOFT60753.2023.10425447.
P. Kedia, P. Soni, P. Gupta, R. Pillai and A. Chaudhary,
"ConvXGDFU - Ensemble Learning Techniques for
Diabetic Foot Ulcer Detection," 2022 4th International
Conference on Advances in Computing,
Communication Control and Networking (ICAC3N),
Greater Noida, India, 2022, pp. 1551-1557, doi:
10.1109/ICAC3N56670.2022.10074466.
S. Sudha and R. S. Sabeenian, "ALEXNet Optimized Cross
layer Convolution Neural Network Algorithm for
Efficient Diabetic Foot Ulcer Identification," 2023
International Conference on Integrated Intelligence and
Communication Systems (ICIICS), Kalaburagi,
India, 2023, pp. 1- 6, doi: 10.1109/ICIICS59993.2023.
10421606
S. G, B. Avinash, M. Shabbir Alam, M. Vyankatesh
Ghamande, K. K. Gupta and R. Rastogi, "Automatic
Detection of Infection in Diabetic Foot Ulcer Images
Using Improved-CNN-SVM Approach," 2023 7th
International Conference on Electronics,
Communication and Aerospace Technology (ICECA),
Coimbatore, India, 2023, pp. 505-510,
T. K. Vaiyapuri and R. Valli, "Segmentation of Diabetic
Foot-Ulcer from Tsallis and Aquila-Optmizer
Enhanced Images," 2023 International Conference on
System, Computation, Automation and Networking
(ICSCAN), PUDUCHERRY, India, 2023, pp. 1-5,
Z. Liu, J. John and E. Agu, "Diabetic Foot Ulcer Ischemia
and Infection Classification Using EfficientNet Deep
Learning Models," in IEEE Open Journal of
Engineering in Medicine and Biology, vol. 3, pp. 189-
201, 2022, doi: 10.1109/OJEMB.2022.3219725.
ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,
COMMUNICATION, AND COMPUTING TECHNOLOGIES
214
