AI‐Enhanced System for Diabetic Foot Ulcer Localization and Classification Using Multi‐Scale Neural CNN Model

ABSTRACT

Background and Aims

Diabetic foot ulcers (DFUs), which heal slowly because of inadequate blood supply, are among the serious infections and chronic foot problems that can result from diabetes mellitus (DM). The primary consequence of DFU, which can result in amputation if left untreated.

Methods

The paper proposes an intelligent and automated approach for classifying foot images as either healthy or DFU images. At the initial stage, the proposed system presents a novel dataset containing 5500 foot images collected from diverse individuals with healthy and DFU conditions. The preprocessing stage involves three steps: the Region of Interest (ROI) method removes unnecessary portions of the foot images, the RGB images are converted to grayscale, and Non‐Local Means (NLM) filtering is applied to reduce noise and remove unwanted information. Two neural feature extractors, ResNet50 and Faster R‐CNN, are used to independently extract features from the foot images, and the resulting feature vectors are integrated into a single fused feature. The softmax function of Faster RCNN method classifies DFU or normal image using fused features, and bounding box method regressor function localize the ulcer region from DFU image.

Results

Compared with relevant state‐of‐the‐art methods, the proposed Faster RCNN‐based deep learning approach with feature fusion demonstrates superior performance in DFU recognition, achieving a testing accuracy of 99.85%, specificity of 99.37%, and precision of 99.50%. With a detection accuracy of 98.83%, this fusion‐based approach demonstrated competitive performance through the use of a generalization validation mechanism.

Conclusion

For automated diabetic foot ulcer identification and localization, the suggested fusion‐based Faster R‐CNN framework shows very accurate and dependable performance. These results point to its significant potential as a helpful clinical decision‐making tool for the early identification and treatment of DFU.

1. Introduction

An unusually high blood sugar level is a hallmark of diabetes mellitus. This chronic illness can cause major and occasionally fatal consequences, such as lower limb amputations, cardiovascular problems, eyesight loss, and kidney impairment [1]. According to the International Diabetes Federation, around 9.3% of people worldwide have diabetes, and by 2030, that percentage is expected to rise to 10.2% [2]. There were only 151 million diabetics globally in 2000; by 2014, that number had risen to about 422 million, and by 2021, it had reached over 537 million [3]. Between 2000 and 2021, the prevalence of diabetes among persons over the age of 18 increased by 10.5%. According to Table 1, there will be 630 million people with diabetes by the end of 2035. Neuropathy, especially in the feet, is a major consequence of diabetes that can result in incurable infections. Due to decreased blood circulation, diabetics frequently have trouble healing foot ulcers [4], which can worsen infections and eventually require amputation. Diabetic foot ulcers (DFUs) also have a very high recurrence rate, with around 40% occurring after the first year and 60% occurring within 3 years [5]. Due to insufficient detection and treatment of DFUs, about 1 million diabetics are amputated each year in the United States alone [6]. In addition to negatively impacting quality of life, such amputation wounds are also prone to complications [7]. Advanced DFU patients are challenging and expensive to treat and care for. Every year, over a million diabetic patients who are at high risk of developing diabetes will lose a portion of their foot due to a lack of resources and a shortage of specialists in the treatment of diabetic foot ulcers. It has been shown that one diabetic foot is operated on every 20 s [8]. Figure 1a–d depicts a healthy, normal foot, while Figure 1e–h depicts diabetes patients' foot ulcers. Careful screening and documentation are necessary for the effective treatment of DFU. Therefore, finding a trustworthy method for the early diagnosis and routine screening of DFU is crucial in order to treat them promptly and affordably before moving on to the next step.

Table 1.

Statistics and projections on the incidence of diabetes and diabetic foot ulcers in 2024.

At the glance	2000	2024	2030	2045
Global population of adults aged 19–79 (in billions)	3.2	7.8	8.5	9.3
Number of diabetics people (in millions)	153	547	650	790
Diabetes incidence (percentage)	4.6%	11%	12.3%	13.2%
15% of people have diabetic foot ulcers (in millions)	22.65	81.55	97.45	118.45
Number of diabetics people in Bangladesh, expressed in millions	22.7	45.2	90	115.9

Figure 1.

Sample of diabetic foot images. (a–d) Normal foot images and (e–h) abnormal images called diabetics foot ulcer which is caused by diabetics.

One of the challenges in managing diabetes mellitus is the early detection and prompt intervention of DFU. Neurological processes may be compromised by a prolonged, persistent hyperglycemic condition, and autonomic neuropathy may result in decreased sweat production, which can cause dry, chapped skin and deep skin structures that are vulnerable to microbial invasion and colonization. Furthermore, disorders in lipid metabolism may encourage atherosclerosis in individuals with type 2 diabetes, resulting in artery constriction and a decreased blood flow to the foot. Foot sensation is reduced by the aforementioned pathogenic processes, and DFU development is accelerated by the unnoticed trauma. Additionally, immunological dysregulation and chronic inflammatory stress impair the body's immune system, which causes delayed wound healing and recurring infections, ultimately leading to amputation and even death [9]. As a result, DFU management and early identification are currently urgent issues.

Timely diagnosis of DFUs is crucial. A multicenter study in Bangladesh showed that delays in diagnosis are common, which can hinder effective treatment by allowing foot disease to progress and creating missed opportunities for intervention. Considering the uneven distribution of medical resources, disparities in healthcare quality, and shortage of medical personnel in Bangladesh [10], it is particularly important to develop a DFU risk warning system and a chronic disease management model that are simple to use and easy to implement widely.

The development and efficacy of deep learning (DL) frameworks and artificial neural networks (ANNs) have led to the rise in popularity of AI‐based computer‐assisted diagnosis (CAD) for a variety of ailments in recent years. Applications using artificial intelligence (AI) are an essential tool for helping overburdened medical professionals advocate for improved procedures. By providing decision support at the point of treatment with prompt and accurate identification of adverse changes in the course of wound healing, they automate repetitive processes. Since 2015, many attempts have been made to diagnose DFU using methods based on artificial intelligence [11]. DFU pictures were analyzed simultaneously using both conventional machine learning (ML) and computer vision (CV) approaches [12]. Researchers at DFU also concentrated on the use of DL in DFU due to recent developments in DL approaches in the CV area; for example, Goyal et al. [13] suggested automatic segmentation in 2017. Since then, some studies have attempted to use planar thermograms to diagnose DFU by looking for hot spots that might be a sign of inflammation or tissue damage [14, 15]. To identify DFU, several academics have conducted extensive study on the classification, detection, and segmentation of foot pictures. However, early detection of DFU cases continues to be a difficult challenge due to the absence of big and appropriately annotated datasets. On DFUC2020, one of the most widely used datasets in this field, the researchers were only able to detect DFU instances with a mean average precision (mAP) of 0.74 [16].

Transfer learning involves reusing a pre‐trained CNN model and is widely used in deep learning for training networks on small datasets [17]. While effective for many tasks, traditional transfer learning struggles with large real‐world datasets, such as those in the medical domain. To overcome this, Multi‐Scale Transfer Learning (MTL) employs multiple pre‐trained CNN models to process data at different scales, capturing features at various levels of detail [18]. The features from each model are then combined into a comprehensive representation, making MTL particularly effective for feature extraction in complex datasets.

The majority of earlier approaches in this subject have focused on creating new capture techniques, but they are not sensitive enough for medical diagnosis, which is important since failing to detect positive instances might have dire repercussions. Other significant research gaps in this area include the following: (i) the localization of DFU has been the subject of very few studies; (ii) ensemble classifiers; (iii) the use of external datasets to evaluate the previous work technique is extremely limited; (iv) the majority of the prior work methodologies have a high level of computational complexity; and (iv) DNN‐based approaches have not been investigated in DFU identification and localization. In this study, we employ a fusion‐based deep neural network (DNN) to achieve high sensitivity in detecting DFU cases, which is critical for patient care. Our bounding box detection approach assists clinical experts in accurately localizing DFU regions, reducing their workload. The proposed AI framework supports early diagnosis, improved treatment planning, minimized complications, and enhanced patient care, while also enabling remote monitoring of patients at home.

The objectives of this work are to apply Faster RCNN techniques for the detection of diabetic foot ulcers (DFUs) and to conduct a comparative analysis of these models [19]. Using a hybrid neural approach, the method preprocesses images, extracts features with neural networks, and integrates them into a single feature vector, enabling the Faster RCNN model to automatically classify images as DFU or normal. By leveraging breadth, depth, and resolution, the model builds an effective prediction framework for DFU diagnosis. Applied to a diabetic foot ulcer dataset, the proposed model facilitates early detection, accurate diagnosis, and improved prognosis [20]. The key novelties of this proposed model are:

• a.Dataset Development: A novel dataset of 5055 foot images was created by combining two publicly available datasets, including both healthy and diabetic foot ulcer (DFU) images from diverse individuals. All images were manually annotated by a DFU specialist.
• b.Preprocessing Stage: To improve the classification result, noise was removed from the foot pictures using a non‐local means (NLM) filtering method.
• c.Fusion‐based Deep Learning Framework: We propose a hybrid model combining ResNet50 and Faster RCNN for feature extraction and classification, improving DFU detection accuracy.
• d.Localization of DFU Regions: The Faster RCNN bounding box approach enables precise localization of ulcers, assisting clinical experts in diagnosis.
• e.High Performance: The proposed method achieves superior detection performance compared to existing approaches, demonstrating its potential for early diagnosis, treatment planning, and remote patient monitoring.

There are five sections in the main body of the paper. The literature study and research conducted by various authors on the principles of diabetic foot ulcer identification are examined in Section 2. Section 3 discusses the method, the preprocessing process, the description of the image collection, and Faster RCNN models. Section 4 discusses the execution and analysis of the results. Section 5 concludes by making inferences from this detection mechanism.

2. Literature Survey

A literature review is carried out to learn about the current works. The main points of the papers pertaining to the suggested work are summarized here.

In order to distinguish between normal and pathological DFU skin from foot photos, the author [21] proposed a unique model called DFU SPNet that is built on stacked parallel convolution layers. Three blocks of parallel convolution layers, each with a range of kernel sizes to extract both local and global information, make up the DFU SPNet model. In order to get an area under the curve (AUC) score of 97.4%, the study additionally concentrated on investigating various optimizers and learning rate combinations. DFU QUTNet is a new network introduced by author [22] for the automatic categorization of DFU pictures. In contrast to other contemporary networks, this one was constructed to expand the network's width while maintaining a respectably high depth. By doing this, gradient propagation is aided and the intricacy of including additional layers into traditional CNN networks is avoided. However, their study mostly treats the topic as a classification problem, which restricts its ability to precisely identify ulcers' locations.

The literature on artificial intelligence (AI) assisted DFU monitoring techniques was thoroughly examined by [23], who also discussed the advantages of these techniques, as well as the challenges of incorporating them into a practical and trustworthy framework for effective remote patient management. The imaging techniques and related optical sensors used to identify diabetic foot ulcers were covered in this work. The analysis takes into account the physiology of the patient, as well as the characteristics of the sensors. The data source provided support for various monitoring strategies, which limits the use of AI technologies [24].

By accurately locating ulcer spots, the research [25] proposed a deep learning‐based method for identifying DFUs from photos of patients' feet. The approach combines predictions from cutting‐edge object identification models using novel model ensemble approaches, such as weighted bounding box fusion (WBF), Soft‐NMS, and non‐maximum suppression (NMS). When merged as an ensemble, the performances of the various state‐of‐the‐art model architectures employed in this study enhance one another and produce more enhanced and broadly applicable outcomes. On the DFUC2020 dataset, our WBF‐based method that combines YOLOv8m with FRCNN‐ResNet101 obtains a mean average accuracy (mAP) score of 86.4% at the IoU threshold of 0.5, which is a remarkable 12.4% improvement over the previous benchmark. The IEEE DataPort Diabetic Foot dataset, which has shown strong and consistent model performance in the qualitative analysis, is another dataset we externally validate.

A comprehensive investigation on real‐time DFU localization for mobile devices was carried out by Goyal et al. (2019) [26]. On a variety of deep learning models, such as SSD MobileNet, Faster RCNN Inception‐v2, and RFCN‐ResNet10, the study employed two‐tier transfer learning. With a mAP of 0.918, Faster RCNN‐Inceptionv2 achieved the highest score. The model was then put into practice for real‐time detection using an NVIDIA Jetson TX2 module and an Android app. No further measures were taken to combine the predictions to increase accuracy once it was discovered that the models employed in this study were unable to reliably forecast tiny ulcer spots.

In order to improve the interpretability of deep learning models for DFU labeling and localization and guarantee clinical relevance, this study [27] presents the DFU_XAI framework. The framework uses interpretability approaches, such as SHAP, LIME, and Grad‐CAM, to analyze six sophisticated models: Xception, DenseNet121, ResNet50, InceptionV3, MobileNetV2, and Siamese Neural Network (SNN). Among these, the SNN model excelled with 98.76% accuracy, 99.3% precision, 97.7% recall, 98.5% F1‐score, and 98.6% AUC. Grad‐CAM heat maps provided doctors with accurate and aesthetically pleasing information by successfully identifying ulcer spots. By incorporating explainability into AI‐driven healthcare, the DFU_XAI framework improves clinical environments' usability and trustworthiness.

The author [28] proposed a transfer learning‐based automated method for detecting and categorizing DFU. DFU is usually divided into two difficult‐to‐visualize states: ischemia and infection. For autonomous DFU identification, we assess how well pretrained Deep Convolutional Neural Network (DCNN) models perform. EfficientNetB0, DenseNet121, ResNet101, VGG16, MobileNetV2, InceptionV3, and InceptionResNetV2 are the seven models that are compared. To reduce overfitting, we also suggest E‐DFu‐Net, a unique model based on pre‐existing architectures. According to experimental data, E‐DFu‐Net performs very well, classifying infections with 92% accuracy and ischemia with 97% accuracy.

This study [29] presents an intelligent automated system for classifying diabetic retinopathy (DR) from retinal fundus images. After preprocessing, discriminative features are extracted using HOG, Shearlet transform, and Faster RCNN, and fused into a single feature vector. A Faster RCNN based classifier is then employed to identify DR and lesion regions. Experiments on three public datasets demonstrate competitive performance, achieving 98.58% accuracy, 97.12% specificity, and 95.72% sensitivity, outperforming existing methods.

The author [30] proposed a novel Convolutional Neural Network (CNN) architecture, termed HCNNet, by integrating hybridized blocks from inception, residual, and dense modules, along with strategically placed Squeeze‐and‐Excitation (SE) blocks and intermediate transition layers. The network is trained using multiple optimizers and learning rate configurations to optimize performance, achieving a high Area Under the ROC Curve (AUC) of 0.999 for ischemia identification.

In this research [31], recent developments brought about by several research initiatives have produced intelligent healthcare monitoring systems that address the security, scalability, and affordability issues facing the healthcare sector. These systems use machine learning (ML) and the Internet of Things (IoT) to change the way of life of the present generation and preserve health for future generations. However, there are many obstacles to overcome before smart healthcare can be achieved using ML and IoT devices, especially when it comes to diagnostic accuracy and privacy protection. This chapter addresses the difficulties encountered in using IoT and ML in smart healthcare while offering suggestions for enhancing the system.

The healthcare sector is undergoing a major transformation through the integration of Artificial Intelligence (AI) and blockchain technology. While AI enables predictive diagnostics, fraud detection, and personalized treatment, blockchain ensures data security, integrity, and transparent access control. Together, these technologies improve decision‐making, optimize resource utilization, and enhance trust in healthcare systems. This paper presents a comprehensive review of blockchain‐based healthcare solutions and examines their integration with AI, highlighting current frameworks, challenges, emerging trends, and future research directions [32].

In this study [33], patients were classified using AI‐based algorithms after extracting features from phonocardiogram (PCG) signals. The XGBoost algorithm outperformed other methods in detecting cardiovascular diseases, achieving specificity, sensitivity, and accuracy of 99% ± 1.93%, 98% ± 2.76%, and 99% ± 1.78%, respectively. The proposed method supports fast and accurate diagnosis and has potential for future extension to heart valve disease detection.

The study [34] proposes a human–computer communication system to assist hearing‐impaired individuals in medical clinics and banks. A Turkish Sign Language dataset, Bosphorus Sign, was developed with input from sign language experts and domain specialists. Using a subset of this dataset, a prototype system called HospiSign was designed to guide deaf users through clinic visits via a tree‐based interaction model. The system recognizes sign responses using hand position, shape, movement, and chest‐area features, with temporal variations modeled using dynamic time warping and temporal templates. Classification is performed using k‐Nearest Neighbors and Random Forest algorithms, and experiments demonstrate the effectiveness of the proposed framework.

The proposed approach [35] addresses intermittent clients using DFU images and applies data augmentation to mitigate class imbalance during local training. A novel CNN architecture, ResKNet (K = 4), is introduced for client‐side learning, featuring residual blocks with 2D convolution, batch normalization, LeakyReLU activation, and skip connections. The framework is evaluated across varying client numbers and dataset sizes, demonstrating its scalability and effectiveness. It enables secure data sharing using consumer devices and supports integration with smartphones and wearables, facilitating collaborative and robust DFU diagnosis across medical institutions.

According to the literature review, patients' lack of knowledge about the repercussions of diabetes is the main cause of diabetic complications. Several sensors [36] and CNN models were used by the researchers to analyze the DFU detection; however, only one or two of the image's width, depth, or resolution were taken into account. Therefore, all three characteristics are taken into account for the prediction model for diabetic foot ulcer identification in this deep learning trial model, EfficientNet. By accurately distinguishing between a diabetic foot and a normal foot, it can monitor the severity of the foot ulcer and provide preventative measures. Table 2 summarizes the recent AI‐based approaches for diabetic foot ulcer (DFU) detection and related healthcare applications, including methodologies, datasets, key findings, and identified limitations.

Table 2.

Summary of AI‐based methods for DFU detection, datasets, results, and limitations.

Ref.	Methodology	Dataset	Key findings/results	Limitations and gaps
[21]	DFU‐SPNet with stacked parallel CNN layers	DFU image dataset	Achieved AUC of 97.4% for DFU classification	No ulcer localization; classification only
[22]	DFU‐QUTNet (wide CNN architecture)	DFU image dataset	Improved gradient flow and classification accuracy	Unable to precisely localize ulcer regions
[23]	Optical imaging & sensor‐based analysis	Sensor‐based DFU data	Effective physiological and sensor‐based monitoring	Limited use of AI techniques
[24]	Survey of AI‐assisted DFU monitoring	Multiple sources	Identified benefits and challenges of AI‐based DFU systems	No experimental framework
[25]	Ensemble DL (YOLOv8 + Faster R‐CNN with WBF)	DFUC2020, IEEE DataPort DFU	mAP 86.4% (12.4% improvement over baseline)	High computational cost
[26]	Transfer learning with Faster R‐CNN & SSD	DFU image dataset	Achieved mAP of 0.918; real‐time mobile deployment	Poor detection of small ulcers
[27]	Explainable AI (Grad‐CAM, SHAP, LIME)	DFU image dataset	SNN achieved 98.76% accuracy with interpretability	Increased computational complexity
[28]	Transfer learning (E‐DFu‐Net)	DFU image dataset	92% infection and 97% ischemia classification accuracy	No DFU region localization
[29]	Feature fusion + Faster R‐CNN	Public DR datasets	Accuracy 98.58% for DR detection	Focused on retina images, not DFU
[30]	HCNNet (Hybrid CNN with SE blocks)	DFU ischemia dataset	AUC of 0.999 for ischemia identification	Limited to ischemia detection
[31]	IoT + ML survey	Multiple healthcare systems	Highlighted challenges in smart healthcare	No practical validation
[32]	AI + Blockchain integration	Healthcare systems	Improved security, transparency, and efficiency	Not image‐based diagnosis
[33]	XGBoost on PCG signals	PCG cardiovascular dataset	Accuracy 99% ± 1.78%	Signal‐based, not imaging
[34]	HCI‐based sign language recognition	Bosphorus Sign dataset	Effective communication for hearing impaired	Not disease‐detection focused
[35]	Federated learning with ResKNet	DFU image dataset	Secure, scalable DFU diagnosis	Dependent on client participation

3. Proposed Model Framework

Detecting Diabetic Foot Ulcer (DFU) from diabetic patient foot images is a four‐step process: preprocessing, feature extraction, classification, and Ulcer region segmentation and Localization between DFU and normal images. The dataset's photos were gathered, properly classified, and combined into a single dataset in the first step. Preprocessing included filtering the original picture using the Non‐Local Means (NLM) approach, converting the RGB images in the original dataset to grayscale photos, and clipping undesired areas from the original images. Following annotation, distinctive features were extracted from the DFU images using the Faster RCNN technique. Furthermore, another feature was extracted using the ResNet50 CNN model, and the two neural feature vectors were then combined. Lastly, the DFU anomalies within the photos were located, and the DFU and normal images were classified using a Faster RCNN model. Figure 2 shows the architecture of the recommended strategy.

Figure 2.

The suggested Faster‐RCNN architecture for DFU classification and localization is shown in this framework.

3.1. Dataset Collection

There are several publicly accessible datasets for identifying normal pictures and DFU. They are widely used for training, testing, and validating various machine learning algorithms, as well as comparing one system's performance to that of others. The datasets were categorized as either normal foot pictures or DFU to serve as a standard for assessing the performance of the intelligent system. Two common datasets were utilized in this study to detect DFU.

• 1.DFU Dataset 1: In this dataset [37], which consists of four folders that are mentioned in the Kaggle online data repositories. This dataset was extracted from the diabetes department of the Nasiriyah Hospital in southern Iraq. All relevant patients and individuals provided written consent and ethical approval, since medical picture data contains sensitive patient information. These samples were taken at different angles and luminosities using Samsung Galaxy and iPad smartphones. Out of these four directories, the author chose the patch folder, which had 1055 skin patches in total. Of these patches, 543 were categorized as normal (healthy skin), while 512 were found to be abnormal (ulcers).
• 2.DFU Dataset 2: Enhancing the precision of DFU identification in practical contexts was the aim of the Diabetic Foot Ulcer Competition 2020 (DFUC 2020) dataset [38]. Foot photos with DFUs that were gathered from Lancashire Teaching Hospitals made up the dataset. Three digital cameras—the Kodak DX4530, Nikon D3300, and Nikon COOLPIX P100 were used to take the pictures, and close‐ups of the foot were shot without the use of zoom or macro features. The dataset includes 4000 high‐resolution DFU pictures, ranging in size from 433 × 289 pixels to 1184 × 1456 pixels, that were taken with many cameras. The DFU is represented in 2250 of the dataset's 4000 photos, with the other 1750 images being normal. The public has access to the training image ground facts.

After the two datasets were combined into a single dataset, 5055 DFU images were separated into 2762 DFU images and 2293 normal images. Every image had a distinct dimension, ranging from 433 × 289 to 1504 × 1456 pixels, and was in RGB format. Figure 3 displays exemplary photos for both the normal and DFU categories. The areas that are highlighted indicate that the picture is broken and that the DFU with foot ulcer has been diagnosed.

Figure 3.

Shown sample dataset of normal and DFU image.

3.2. Preprocessing Stage

The data preprocessing stage in our study consists of three main steps designed to enhance image quality and prepare the dataset for robust DFU classification: Region of Interest (ROI) Extraction, Grayscale Conversion, and Noise Reduction with NLM Filtering. Unnecessary text, annotations, and background areas were removed from each image, and the DFU region was cropped to 224 × 224 pixels. Next, Python libraries were used to convert the RGB photos to grayscale. A Non‐Local Means (NLM) denoising method has been applied to remove high‐frequency noise while preserving important structural details in the foot images. Then the researcher divided two parts, named training and testing. The preprocessing stage is a very powerful techniques that fix overfitting problems using a balanced dataset, increases the model's resilience, and improves the proposed framework's performance. Figure 4 illustrates the output of the preprocessing stage.

Figure 4.

Date shown after performing every step of the preprocessing phase.

3.2.1. Non‐Local Means Filtering Method

The main goal of filtering techniques is to isolate the most important parts of the image and remove noise. To obtain the appropriate features or essential details of any noisy image datasets, the filtering methods are most suitable. DFU images contain high‐frequency noise, which is often eliminated using filtration. It is crucial that the original picture, and in particular its features, sustain as much of its structural integrity as possible during the noise reduction process. On the basis of this, several approaches to noise reduction have been put forth. For this objective, we employed the recently developed Non‐Local Means (NLM) denoising approach in this work. Unlike traditional filters, NLM effectively removes high‐frequency noise while preserving important structural and textural features of DFU images, which are critical for accurate feature extraction and classification. Its ability to consider the similarity of pixels across the entire image, rather than just local neighborhoods, ensures better retention of fine details and improves the model's overall performance in detecting and localizing ulcers.

The NLM filter is an upgraded version of the Yaroslavsky filter [39] that determines the true number of pixels being processed by nonlocally averaging related pixels, or pixels with a closer brightness level. The primary benefit of the NLM approach over this one is that, in the presence of noise, its similarity criteria are more stable since, in addition to comparing the intensity levels of the pixels, a neighbor of each pixel also influences the degree of similarity. Most noise models can be effectively reduced using the NLM approach, particularly if the noise can be dispersed collectively. The NLM approach is based on assigning a weight to each pixel in the picture depending on how similar its neighbors are to the pixel being processed. Put another way, the higher the weight given to the image pixel neighbors, the more similar they are to each other. The total weight discovered from the other pixels is used to compute the number of pixels being processed. Equation 1 illustrates the weighted Euclidean principle with the Gaussian kernel, which is the neighborhood criteria similarity in the NLM technique.

$$d={\Vert v\left(n_i\right)-v\left(n_j\right)\Vert }_{2\rho }^2=\sqrt{\sum _{k=1}^N{\omega }_{\mathit{ik}}}{(x_{\mathit{ik}}-x_{\mathit{jk}})}^2$$ (1)

where ${\Vert ·\Vert }_{2\rho }^2$ stands for the weighted Euclidean distance operator with Gaussian kernel, $v\left(n_i\right)$ denotes the neighborhood vector of the pixel under process and the neighborhood vector of neighboring pixels. Put another way, the influence of the pixels reduces as one moves past the central pixel, which has a larger value for determining the similarity of neighborhoods.

$$W(M_i,M_j)=\frac{1}{Z_i}\text{exp}\left(-\frac{{\Vert v\left(n_i\right)-v\left(n_j\right)\Vert }_{2\rho }^2}{h^2}\right)$$ (2)

where

$$Z_i=\sum _j\text{exp}\left(-\frac{{\Vert v\left(n_i\right)-v\left(n_j\right)\Vert }_{2\rho }^2}{h^2}\right)$$ (3)

where $Z_i$ is a normalization parameter that ensures the weights used add up to 1. $h$ defines the key parameter of the NLM that affects the filtering strength. The parameter $h$ controls the degree of smoothing in the Non‐Local Means (NLM) filter. If $h$ is picked small, the value of the filtering in the picture is tiny, and the noise effect has been not eliminated adequately, but a big value for $h$ causes an overfiltering for the image, and the reconstructed image is fully blurred and devoid of fine structural features. After testing several values, we found that h = 10 provided the best balance between noise removal and preservation of structural details in the foot images. The NLM filter's final equation with calculated weighted coefficients may be expressed as follows:

$$\mathrm{NLM}\left(M_i\right)=\sum _{j\in \mathrm{\Omega }}W(M_i,M_j)M_j$$ (4)

where $M_i$ is the pixel being denoised, $M_j$ represents pixels within the search window Ω, and $W(M_i,M_j)$ is the normalized similarity weight between pixels $M_i\mathrm{and}{M}_j$. Therefore, the weighting process is limited to the region known as the search window that surrounds each processed pixel. As previously mentioned, NLM is a parameter filter that has three parameters: smoothing parameter ($h$), search window radius, and similarity window radius. An example of noise reduction in this scenario is shown in Figure 5.

Figure 5.

Represent the filtered image after applying non‐local means filtering method.

3.3. Faster RCNN Feature Extractor

A deep, fully linked network, a region proposal network (RPN), and a fast RCNN detector are the two components that make up the Faster RCNN classifier. Finding the ideal areas of different sizes and proportions is the goal of the region proposal network. Using a region proposal technique, quicker RCNN constructs the collections of regions. The regional proposal network, which is an extra CNN in Faster RCNN, is used to get regional suggestions. After receiving the feature map as input, the proposal network in the training region generates region suggestions. The ROI pooling layer receives these suggestions and processes them further [40].

There are four pooling layers and 13 convolutional and ReLU layers in the Faster RCNN. These convolutional layers are used by the Faster RCNN network to create the input image's feature map, which is then sent to the RPN module and other layers. Anchors and bounding box regression offsets are constructed using the 3 × 3 fully convolutional layers that comprise the RPN module.

To ascertain whether the calculated anchors are in the front or background, the Softmax function is used. At the last step, the generated anchors and bounding boxes are used to calculate the object recommendations. Proposal feature maps are created by this layer and sent to other relevant network levels. Lastly, the diabetes foot ulcer locations are determined by the classification stage. It works by using the output of the ROI pooling layer. Bounding box regression is used to illustrate the detected test box's final position.

3.4. ResNet50 Feature Extractor

ResNet50 is a potent picture classification model that produces cutting‐edge outcomes when trained on enormous data sets. The use of residual connections, which enable the network to learn a set of residual functions that map the input to the intended output, is one of its main advances. The network has 50 layers deep [41]. The identity block, convolutional block, fully connected layers, and convolutional layers make up the four key components of the ResNet50 architecture. The identity block and convolutional block are in charge of processing and converting the features that the convolutional layers have extracted from the input image. Lastly, the ultimate categorization is determined using the fully connected layers. Multiple convolutional layers are followed by batch normalization and ReLU activation in the convolutional layers of ResNet50 [42]. These layers are in charge of taking characteristics like edges, textures, and forms out of the input picture. Max pooling layers, which decrease the spatial dimensions of the feature maps while maintaining the most significant features, come after the convolutional layers. ResNet50 mitigates the vanishing gradient problem by employing identity‐based skip connections that facilitate direct gradient propagation across layers during back‐propagation. This architecture enables the network to effectively train deeper layers and learn richer, more complex feature representations without sacrificing performance.

3.5. Implementing Multi‐Scale CNN (MS‐CNN) Model

In order to construct the MS‐CNN model, the DFU pictures with a size of 224 × 224 are initially given as input to two machine learning models ResNet50 and Faster RCNN. This allows each model to extract the characteristics of the images individually. The GlobalAveragePooling2D layer of ResNet50 CNN model provides a feature map sized$1\times 1\times 2048$. Then the ROI pooling layer of Faster RCNN model provides a feature map sized $1\times 1\times 512$. The concatenate layer is then used to merge these separate feature vectors into a single vector. Finally, the fully connected layer of faster RCNN classify normal and DFU images [43]. Another fully connected layer is bounding box regression that localized the ulcer region from the DFU images. The MS‐CNN model for DFU picture classification is shown in Figure 6. Batch normalization speeds up convergence by stabilizing intermediate feature distributions, reducing internal covariate shift, and enabling higher learning rates. ReLU activations further enhance convergence by preventing saturation and alleviating the vanishing gradient problem, allowing efficient gradient flow in deep networks. Feature fusion is mainly intended to improve model accuracy rather than convergence speed by integrating complementary features from networks, such as ResNet50 and Faster R‐CNN, thereby producing richer representations that enhance classification and localization performance.

Figure 6.

Illustrates the implementation of multi‐scale CNN model.

3.6. Faster RCNN Classification and Segment Ulcer Region

The Faster RCNN method classifies the normal and DFU images from the input image by using fusion feature map. The Convolutional Backbone extracts hierarchical features from input foot images. The Region Proposal Network (RPN) generates candidate object regions (potential DFU locations) by predicting objectless scores and bounding boxes. The ROI Pooling layer converts variable‐sized region proposals into fixed‐size feature maps for classification and localization [44]. Finally, the fully connected layers classify each region as DFU or normal and refine bounding box coordinates. Once the DFU retina pictures have been classified, the algorithm locates the afflicted area. The bounding box regression is used to show the final location of the determined test box. The backdrop and surrounding areas are utilized as a negative example for foot ulcer region sign localization, while the affected regions are used as a positive example. Intersection over Union (IoU), a threshold value of 0.3, is used to classify the overlapping region as background or detrimental. The area is regarded as negative if the IoU value is less than 0.3. Likewise, areas that have an IoU value higher than 0.7 are considered to be DFU lesions. The Faster RCNN method was applied to the localization of the DFU lesions. The bounding box regression can restrict localization accuracy under varying conditions, including differences in ulcer size, shape, and lighting [45]. To overcome this limitation, our approach employs feature fusion from ResNet50 and Faster R‐CNN, which enriches the feature representation and helps mitigate this issue by capturing more comprehensive contextual information. The dataset was subjected to the localization outcome of the Faster RCNN, as seen in Figure 7.

Figure 7.

Localized ulcer region using the Faster RCNN regression layer.

4. Experimental Results and Perspectives

4.1. Evaluation Criterion

The performance of the suggested framework was evaluated using a number of statistical measures, such as accuracy (ACC), specificity (SP), sensitivity (SE), precision (PRE), recall (REC), F1‐score (FS), Kappa statistic, and MCC. These parameters were calculated using the values of false negative (FN), true positive (TP), true negative (TN), and false positive (FP) in the confusion matrix. TP stands for normal foot pictures that were accurately detected throughout the evaluation of the proposed framework. TN is the exact identification of a DFU picture. On the other hand, FN denotes incorrect DFU picture identification, whereas FP denotes incorrect normal foot image identification. The proposed framework's performance metrics are as follows in Equations (5)–(12):

$$\mathrm{Accuracy}\left(\mathrm{ACC}\right)=\frac{\mathrm{TP}+\mathrm{TN}}{(\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN})}$$ (5)

$$\mathrm{Specificity}\left(\mathrm{SP}\right)=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}$$ (6)

$$\mathrm{Sensitivity}\left(\mathrm{SE}\right)=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

$$\mathrm{Precision}\left(\mathrm{PRE}\right)=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$ (7)

$$\mathrm{F}1\mathrm{Score}\left(\mathrm{FS}\right)=2\times \frac{\mathrm{PRE}\times \mathrm{REC}}{\mathrm{PRE}+\mathrm{REC}}$$ (8)

$$\mathrm{Kappa}=\frac{\mathrm{Total}\mathrm{ACC}-\mathrm{Random}\mathrm{ACC}}{1-\mathrm{Random}\mathrm{ACC}}$$ (9)

$$\mathrm{MCC}=\frac{(\mathrm{TP}\times \mathrm{TN})-(\mathrm{FP}\times \mathrm{FN})}{\sqrt{(\mathrm{TP}+\mathrm{FP})(\mathrm{TP}+\mathrm{FN})(\mathrm{TN}+\mathrm{FP})(\mathrm{TN}+\mathrm{FN})}}$$ (10)

$$\mathrm{Recall}\left(\mathrm{REC}\right)=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$ (11)

$$\mathrm{IoU}=\frac{\mathrm{Area}\mathrm{of}\mathrm{Overlap}}{\mathrm{Area}\mathrm{of}\mathrm{union}}$$ (12)

4.2. Experimental Settings

In order to validate the proposed system developed for automated DFU image recognition. The DFU dataset consists of 5055 images for training, testing, and validation, comprising 2762 DFU images and 2293 normal images. In this work, the data distribution was laminated to solve the problem of data disequilibrium. For training and testing, the proposed dataset was divided into 0.8 and 0.2 halves, respectively. Table 3 displays the data distribution used to create the DFU detection framework.

Table 3.

Data distribution for diabetic's foot ulcer detection system.

Labeled data	Training data	Testing data	Validation data
Normal	1834	459	109
DFU	2209	553	103
Total sample	4043	1012	212

A PC running 64‐bit Windows with an Intel Core i7 CPU operating at 3.00 GHz and 8 GB of RAM was used for all computations. All relevant experiments were carried out using the Python Colab program. The experimental timings were calculated using the GEFORCE RTX 2070 Super GPU combo.

4.3. Filter Method Performance Analysis

The research employed multiple filtering methods, and the filtering effectiveness was assessed based on criteria such as Peak Signal to Noise Ratio (PSNR) [46], Mean Square Error (MSE) [47], and Normalized Cross Correlation (NK) [48]. The MSE, PSNR, and NK formulae are given by Equations (13)–(15). The DFU image influenced by Gaussian noise is subjected to each filter. As seen in Table 4, the recommended filtering method, the NLM filtering method, denoised the picture more effectively than the Adaptive Median Filter (AMF), Antistrophic diffusion filter, Gaussian filter, and Laplacian filter. In terms of MSE, PSNR, and NK, NLM yielded the lowest values, that is, 3.89 and 0.0057. This demonstrated unequivocally that the NLM delivered the highest quality in the picture that was recovered from the noisy image.

$$\mathrm{MSE}=\frac{1}{\mathrm{MN}}\sum _{j=1}^M\sum _{k=1}^N{(X_{j,k}-X_{j,k})}^2$$ (13)

$$\mathrm{PSNR}=10\log \frac{{\left(255\right)}^2}{\mathrm{MSE}}$$ (14)

$$\mathrm{NK}=\sum _{j,k}^M\sum _{j,k}^N{X}_{j,k}-X^{\prime }j,k^{\prime }\sum _{j,k}^M\sum _{j,k}^N{X}_{j,k}^2$$ (15)

Table 4.

Comparative analysis of several filtering method with proposed filtering method.

Several filtering techniques	Filtering method performance measurement metrics
	MSE	PSNR	NK
Adaptive median filter	218.7323	23.5912	0.0934
Antistrophic diffusion filter	13.6420	36.8901	0.0180
Gaussian filter	128.6235	28.4563	0.0890
Laplacian filter	230.3267	27.1065	0.0913
NLM proposed filter	3.8921	39.8901	5.7834e‐4

4.4. Individual and Neural Combined Feature Performances

Initially, the dataset was trained using multiple machine learning models (ResNet50, AlexNet, VGG19, VGG16, EfficientNet, DenseNet121, MobileNet, FasterRCNN, FastRCNN, and U‐Net), among which ResNet50 and Faster R‐CNN demonstrated superior performance. Also, ResNet50 was chosen due to its deep architecture and residual learning capability, which effectively addresses the vanishing gradient problem and enables robust extraction of high‐level semantic features from medical images. On the other hand, Faster R‐CNN was selected for its performance in object detection and localization. It integrates a Region Proposal Network (RPN) with a detection network, allowing accurate identification and localization of ulcer regions within foot images. Therefore, these two models were integrated to leverage their complementary strengths and achieve improved overall results.

Finally, features extracted using several techniques were merged, and the top‐performing feature combination was determined using a quicker RCNN classification methodology. Additionally, the Faster RCNN model's neural features are combined with another ResNet50 neural feature utilizing feature maps. However, compared to the interpolation‐oriented descriptor, concatenating Faster RCNN and ResNet50 features shows better classification results. The results achieved utilizing various fusion approaches and individual feature extractors are compared in Table 5. In comparison to the other double‐fused techniques, the feature vector derived from the fusion of two neural approaches performed better.

Table 5.

Performance analysis of different CNN model with proposed method.

ML model	ACC	SPE	SEN	PRE	Recall	F1‐score	MCC	Kappa	Error rate	mAP	AUC
VGG16	0.895	0.956	0.952	0.969	0.823	0.952	0.794	0.737	0.158	0.65	0.921
VGG19	0.981	0.972	0.978	0.969	0.974	0.978	0.969	0.962	0.017	0.69	0.983
ResNet50	0.989	0.993	0.991	0.975	1.00	0.991	0.991	0.923	0.013	0.78	1.00
Faster RCNN	0.994	0.987	0.965	0.997	0.912	0.965	0.973	0.984	0.024	0.82	0.954
MS‐CNN model,(ResNet50+Faster RCNN)	0.998	0.993	0.991	0.995	0.991	0.993	0.998	0.995	0.009	0.91	0.996

The model was trained for a fixed number of epochs (epochs 35) using the Faster RCNN architecture with fused neural features. Faster RCNN extractors were utilized to merge several CNN‐based feature extractors, and the performance of the fused features was also examined separately. When compared to other CNN feature extraction techniques like VGG19, ResNet50, and VGG16, the greatest performance was obtained by combining the quicker Faster RCNN and ResNet50 features. During training, the softmax function of the Faster RCNN architecture learned to classify images as DFU or normal while simultaneously localizing ulcer regions. Also, the Faster RCNN‐based bounding box regressor function localizes the ulcer region from the classified image. Table 5 illustrates the different CNN models with proposed model performance analysis using several performance metrics (here, ACC = accuracy, SPE = specificity, SEN = sensitivity, PRE = precision, AUC = area under the ROC curve).

4.5. Classification Performance

The outcomes shown in Figure 8 were attained during the suggested model's training and validation. The dataset was separated into training and validation sets following preprocessing. A predetermined number of epochs was used to train the model, and the deep learning framework automatically recorded the classification accuracy and loss values for the training and validation data at the conclusion of each epoch. The accuracy and loss curves seen in Figure 8 were created by plotting these recorded data. As indicated in Table 4, classification was carried out using several CNN classifier models in addition to ResNet50 after the two feature vector data were combined into a single vector. With an accuracy of 99.85% and a specificity of 99.37%, the quicker RCNN model once again outperformed the other Faster RCNN or ResNet50. Various performance accuracy and loss curves for the classification are displayed in Figure 8. For the outliers in the validation datasets, certain validation features could have deviated more than the training features, the ResNet50 validation accuracy curve (Figure 8c) drastically lowers.

Figure 8.

Accuracy vs. loss curve performance of several machine learning models. (a) Proposed model (ResNet50+Faster RCNN) accuracy and loss curve. (b) Accuracy and loss curve of the Faster RCNN model. (c) Accuracy and loss curve of the ResNet50 model.

The confusion matrix produced from the performance measurement of the entire system during the classification stage is displayed in Figure 9. In order to assess its intended outcome during the assessment, the proposed method required annotated test data. The ResNet50 model correctly identified 102 DFU pictures out of 103 photos, while failing to identify 1 DFU image. The Faster RCNN model correctly identified 99 DFU pictures out of 103 photos, while failing to identify 4 DFU images. The proposed model properly identified 107 normal photos out of 109 healthy images; however, it was unable to identify 2 normal images.

Figure 9.

Analysis of confusion matrices between several CNN models and the suggested model. (a) Confusion matrix for the ResNet50 model. (b) Confusion matrix for the Faster RCNN model. (c) Confusion matrix for the proposed model.

Table 6 summarizes the hyperparameter settings used to train the proposed model. Multiple training configurations were evaluated, and the final model was developed using the parameter set that achieved the best overall performance. The optimizer and gradient descent loss functions are two crucial elements for selecting hyperparameters throughout the framework training process. We chose RMSProp as the optimizer function for our system because it combines the best features of the Adam and AdaGrad optimizers to efficiently handle sparse gradients in big datasets. We took into consideration a loss function called binary cross‐entropy because our model is focused on binary classification. Determining an optimal learning rate is critical to minimize the loss function, which is quite tough. All models were trained using the same dataset split and under the same experimental circumstances to guaranty a fair comparison, preprocessing pipeline, and evaluation metrics. Hyperparameters were optimized using the validation set, and the final experiments employed a learning rate of 0.0001, batch size of 32, 35 epochs, the RMSProp optimizer, and binary cross‐entropy loss.

Table 6.

An overview of the suggested model architecture's training parameters.

Serial no.	Training parameter	Value	Serial no.	Training parameter	Value
1	Optimizer	RMSProp	6	Loss function	binary_crossentropy
2	Learning rate	0.0001	7	Factor	0.2
3	Minimum learning rate	1e‐5	8	Patience	5
4	Epochs	35	9	Decay	0.00001
5	Batch size	32	10	Metrics	Accuracy

4.6. Checked Proposed Model Validation Performance

Generalization techniques were used to further validate the system's performance. Generalization is a notion that quantifies a model's capacity to react to new inputs [49]. Once trained, this technique can handle new data and produce accurate projections. The DFU picture quality issues in the previous study had a detrimental effect on the generalization performance. This work used the academic torrents DFU dataset, which was completely different from the training set, for generalization reasons. This model had never seen or heard of this dataset before. Six hundred DFU images were utilized in total, divided into two classes: normal (250 images) and DFU (350 images). Figure 10a shows the accuracy vs. loss curve applying the generalization method, and Figure 10b displays the generalization findings' confusion metrics. Just 5 out of 350 DFU photos and 7 out of 250 normal images had a missed detection. With a generalization accuracy of 98.83%, the DFU images exhibited the lowest percentage of false positives. This proved that the algorithm could reliably classify pictures even when using a whole new set of data.

Figure 10.

Shown in the accuracy vs. loss curve and confusion matrix for the generalization method. (a) Shown in the accuracy vs. loss curve applying the generalization method. (b) Shown in the confusion matrix applying the generalization method.

4.7. Comparative Analysis of Relevant Research

In this study, we proposed an AI‐enhanced architecture for DFU localization and classification using a Multi‐Scale Neural CNN model. The primary objective was to improve the accuracy and robustness of DFU detection by leveraging advanced deep learning architectures. The results obtained from our proposed model demonstrate its better performance as compared to current cutting‐edge techniques.

Deep learning methods have been used for DFU identification and classification in several recent studies, as shown in Table 7. Das et al. [20] proposed DFU SPNet, which achieved 96.40% accuracy. Similarly, Alzubaidi et al. [22] introduced DFU OUTNet, reporting an accuracy of 94.5%. Goyal et al. [26] employed SSD MobileNet, Faster RCNN, and RFCN‐ResNet10, achieving a precision of 91.8%. Meanwhile, F Almufadi N [27] used a combination of Capture Box and Two‐Step SVM Classification, yielding a sensitivity of 73.3% and specificity of 94.6%. In contrast, our proposed model, which utilizes a ResNet50+Faster RCNN architecture trained on the DFU dataset, outperformed these existing methods, achieving an accuracy of 99.80%, specificity of 99.30%, and recall of 99.10%. These findings show a notable increase in robustness and detection accuracy, which lowers false positives and false negatives.

Table 7.

Comparison of proposed model with existing methods for DFU detection.

Author (reference)	Dataset	Methodology	Result
Das SK. [20]	DFU_Net	DFU SPNet (stacked parallel convolution layers)	Accuracy = 96.40%
Alzubaidi et al. [22]	Not specified	DFU QUTNet (wider deep CNN)	Accuracy = 94.5%
Rathore, P.S. [26]	Not specified	SSD MobileNet, Faster RCNN Inception‐v2, RFCN‐ResNet10 (two‐tier transfer learning)	Precision = (91.8%)
F Almufadi N. [27]	Not specified	Capture Box + Two‐Step SVM Classification (superpixels for segmentation)	Sensitivity = 73.3%, Specificity = 94.6%
Proposed model	DFU dataset	ResNet50+Faster RCNN	Accuracy = 99.80% Specificity = 99.30% Recall = 99.10%

4.8. Limitation of the Proposed Model

Although the suggested framework performs well in DFU classification and localization, several limitations remain. First, the use of fixed IoU thresholds in the Faster R‐CNN model may reduce localization accuracy for lesions with significant variability in size and shape. Second, the feature fusion and denoising steps introduce additional computational overhead, which may limit real‐time deployment on low‐resource devices.

5. Conclusion

In this study, we presented an automated and intelligent framework for the classification and localization of DFU using a fusion‐based Faster R‐CNN model. A dataset of 5055 foot photos was used in the investigation, combining DFU and healthy images, which were preprocessed using ROI cropping, grayscale conversion, and NLM denoising to remove noise while preserving key features. Feature extraction was performed using ResNet50 and Faster R‐CNN networks, and the resulting feature vectors were fused into a single representation for classification. The suggested model achieved superior performance, with an accuracy of 99.85%, specificity of 99.37%, and precision of 99.50%, outperforming conventional CNN classifiers. Additionally, the model successfully localized ulcer regions in foot images, demonstrating its potential for clinical applications and early DFU detection. In accordance to these findings from experiments, the Faster RCNN classifier performed the best with mAP, IoU, and needed time values of 0.965, 0.975, and 0.20 s, outperforming ResNet50 and other neural network classification models. The accuracy and loss curves, along with confusion matrices, further confirmed the model's stability, generalization capability, and robustness against dataset variability. Overall, this study highlights the effectiveness of deep learning‐based fusion approaches in improving DFU detection and classification, which could significantly support clinical decision‐making and patient care.

While the proposed framework shows promising results, several avenues remain for future research. First, the model can be extended to include multi‐class classification for different stages or types of DFU, such as ischemia and infection, thereby providing more detailed diagnostic information. Second, integrating additional imaging modalities, such as infrared thermography or hyperspectral imaging, could enhance feature representation and improve prediction accuracy. Finally, the deployment of this model on mobile and edge devices could facilitate real‐time DFU detection in remote or home‐care settings, improving accessibility and patient monitoring.

Author Contributions

Md. Nur‐A‐Alam: conceptualization, investigation, methodology, validation, visualization, software, data curation. Md. Anwar Hussen Wadud: writing – original draft, visualization, validation, formal analysis, resources. Mostofa Kamal Nasir: supervision, data curation, resources, writing – review and editing. Khandaker Mohammad Mohi Uddin: methodology, writing – original draft, supervision, visualization, investigation.

Funding

The authors have nothing to report.

Disclosure

The corresponding author (Khandaker Mohammad Mohi Uddin) affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.

Ethics Statement

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The data that support the findings of this study are openly available in DFU_QUTNet dataset at https://www.kaggle.com/laithjj/diabetic-foot-ulcer-dfu. The datasets used in this study are publicly available. The DFU_QUTNet dataset (Alzubaidi et al., 2020) is available on Kaggle at https://www.kaggle.com/laithjj/diabetic-foot-ulcer-dfu, and the DFUC 2020 dataset (Cassidy et al., 2021) can be accessed via https://doi.org/10.17925/EE.2021.17.1.5. All datasets were used in accordance with their respective data usage policies.