Hybrid Deep Learning Framework for Diabetic Wound Assessment: Integrating Segmentation, Classification, and Explainable AI in Cloud-Based Telemedicine Systems

A.DEEP LEARNING IN MEDICAL IMAGE SEGMENTATION

Deep learning-based segmentation has become a cornerstone of medical image analysis, enabling precise delineation of pathological regions in complex clinical images. Early work introduced fully convolutional networks (FCNs) for pixel-level prediction, paving the way for automated and end-to-end segmentation frameworks [11]. A major breakthrough was the U-Net architecture, which employs an encoder–decoder structure with skip connections to preserve spatial resolution and contextual information. This design has proven particularly effective for segmenting irregular anatomical structures, including DFUs, tumors, and skin lesions.

Building on this foundation, subsequent studies proposed architectural extensions to improve segmentation accuracy and robustness. Residual U-Net (ResUNet) integrates residual connections to enhance gradient propagation and mitigate overfitting in limited medical datasets [12]. Attention U-Net further refines this approach by introducing attention gates that suppress irrelevant background regions and emphasize clinically salient features, making it well suited for complex DFU images with heterogeneous appearance. More advanced models, such as Mask R-CNN and DeepLabV3+, incorporate multi-scale feature extraction and instance-level segmentation, achieving higher IoU and Dice coefficients on challenging datasets acquired under varied lighting conditions and skin tones [13].

Despite these advances, most segmentation studies prioritize pixel-level accuracy metrics, such as IoU and Dice coefficient, without examining how segmentation quality affects downstream clinical tasks. In the context of DFUs, accurate boundary delineation alone is insufficient, as segmentation models do not encode semantic or pathological information related to wound severity. Moreover, reported performance is often based on curated datasets, limiting generalizability to real-world telemedicine images that frequently contain occlusions, illumination artifacts, and diverse wound morphologies. These limitations indicate that, while segmentation models are essential for lesion localization, they are inadequate as standalone solutions for clinically meaningful DFU assessment.

B.CLASSIFICATION NETWORKS FOR CLINICAL IMAGE ANALYSIS

Parallel to advances in segmentation, deep learning has substantially transformed the classification of medical images. Early CNNs demonstrated strong capability in identifying pathological patterns by learning hierarchical spatial and textural representations from labeled data [14]. The introduction of ResNet addressed the vanishing-gradient problem through residual connections, enabling the training of deeper and more expressive networks. Subsequent architectures, including DenseNet and EfficientNet, further improved classification performance by optimizing feature reuse and compound scaling across network depth, width, and resolution.

In the context of DFU assessment, CNN-based classifiers have been applied to predict ulcer severity, infection status, and healing potential from clinical photographs [15]. Although these models often achieve high image-level accuracy, they typically operate on full-frame images without explicitly isolating the wound ROI. As a result, classification performance can be degraded by background noise, confounding visual cues, and irrelevant anatomical features. This limitation is particularly problematic in telemedicine scenarios, where image quality and framing are highly variable.

A more critical methodological limitation is that most DFU classification approaches treat severity grading as a nominal multi-class problem, disregarding the ordinal progression inherent in clinical wound scales such as the Meggitt–Wagner system. This simplification can lead to clinically implausible errors, including large grade jumps (e.g., misclassifying advanced ulcers as mild cases), which may have significant implications for treatment planning and patient outcomes. To address this issue, ordinal regression approaches, including cumulative link models and the Consistent Rank Logits (CORAL) framework, have been proposed to explicitly model ordered relationships between severity levels [16]. These methods reduce large-grade prediction errors and improve clinical interpretability by aligning algorithmic outputs with the progressive nature of disease severity.

Despite these advances, most existing DFU classification models still suffer from two key shortcomings. First, ordinal-aware learning is not consistently integrated into state-of-the-art classification pipelines. Second, many high-performing classifiers operate as black-box systems and do not provide explainable outputs that allow clinicians to understand or validate model predictions. Consequently, despite strong performance in image-level prediction, current DFU classification models remain limited in clinical reliability, interpretability, and real-world applicability.

C.HYBRID DEEP LEARNING ARCHITECTURES

The limitations of single-stage deep learning models have motivated the development of hybrid architectures that integrate segmentation and classification within a unified pipeline [17]. Early hybrid frameworks typically adopted a two-step design in which the lesion region is first segmented and then classified according to disease severity. This “image-to-mask-to-grade” paradigm improves classification accuracy by directing the model’s attention to clinically relevant regions while reducing the influence of surrounding tissue.

Between 2019 and 2024, hybrid architectures evolved to incorporate more advanced segmentation and classification components. Studies employing U-Net or Attention U-Net for ROI extraction, followed by classifiers such as EfficientNet or ResNet, reported significant improvements in accuracy and macro-F1 scores [18]. A key advantage of these modular designs is their flexibility, allowing individual components to be optimized or updated independently without retraining the entire pipeline.

More recent work has examined error propagation within hybrid pipelines, demonstrating that segmentation quality directly affects downstream classification performance. Inaccurate or incomplete segmentation masks can lead to erroneous severity predictions, particularly for advanced ulcers occupying small or irregular regions of the image. These findings underscore the importance of accurate and stable segmentation as a prerequisite for reliable severity grading in hybrid systems.

Despite these advances, existing hybrid approaches continue to exhibit important limitations. Many treat severity grades as nominal categories rather than ordered stages of disease progression, resulting in clinically inconsistent predictions. Moreover, XAI mechanisms are often absent or incorporated only as post hoc visualization tools, limiting clinicians’ ability to understand and validate model decisions [19]. Consequently, while hybrid segmentation–classification pipelines improve grading accuracy by focusing on lesion regions, most existing approaches fail to integrate ordinal-aware learning with explainable mechanisms, leaving a critical gap in transparent and clinically consistent DFU assessment suitable for real-world clinical decision support.

D.XAI AND INTEGRATION WITH TELEMEDICINE

As deep learning models become increasingly complex, the demand for XAI in healthcare has grown substantially. Explainability enhances transparency in model decision-making, enabling clinicians to verify whether algorithmic attention aligns with medically relevant features and established clinical reasoning [20]. Among the most widely adopted visualization techniques, Grad-CAM projects class-specific gradients onto convolutional feature maps to generate heatmaps that highlight image regions most influential in a given prediction. Grad-CAM and its extensions have been successfully applied across various medical imaging tasks to improve interpretability and clinician confidence.

Subsequent methods, such as Score-CAM, refine gradient-based approaches by weighting activation maps using class scores, resulting in more stable and visually coherent explanations. Beyond visualization, XAI techniques have been incorporated into clinical auditing, model validation, and human-in-the-loop decision-making frameworks, reinforcing accountability and trust in automated systems. The growing body of evidence indicates that transparent and interpretable models are essential not only for regulatory compliance but also for facilitating clinical acceptance of AI-assisted diagnostics.

In parallel, advances in telemedicine, cloud computing, and the Internet of Things (IoT) have transformed healthcare delivery by enabling remote monitoring and real-time decision support. Edge- and cloud-based AI systems can now perform near real-time image analysis, supporting wound monitoring, triage, and treatment guidance in geographically dispersed or resource-limited settings [21]. Deep learning pipelines optimized for low-latency inference can be integrated into smartphone or web-based platforms, allowing clinicians to assess patient-submitted images without requiring in-person visits.

Despite these technological advances, significant limitations remain. In telemedicine environments, where clinicians must rely on remote assessments and automated recommendations, AI systems must balance efficiency with transparency. Black-box models that deliver rapid but opaque predictions pose ethical, clinical, and medico-legal risks. Moreover, explainability is often evaluated in isolation and rarely integrated with task-specific constraints such as ordinal severity modeling, which is critical for clinically consistent disease grading. Current systems frequently optimize inference speed or accuracy independently, rather than jointly addressing transparency, reliability, and scalability.

These limitations underscore the need for unified hybrid frameworks that integrate XAI with ordinal-aware severity modeling and efficient deployment strategies. Such frameworks are essential for supporting trustworthy, real-time DFU assessment in telemedicine settings and for aligning AI-assisted diagnostics with clinical workflows and decision-making requirements.

E.SYNTHESIS AND RESEARCH GAP

The reviewed literature demonstrates substantial progress in deep learning-based medical image analysis over the past decade. Notable advancements include improved segmentation accuracy through attention and multi-scale architectures, enhanced classification performance enabled by efficient and ordinal-aware networks, and increasing emphasis on explainability to facilitate clinical adoption [22]. Despite these developments, several critical challenges remain insufficiently addressed in the context of DFU assessment.

First, segmentation and classification are often treated as independent stages, limiting end-to-end interpretability and preventing systematic analysis of how localization quality influences severity grading. Second, many DFU classification systems continue to model severity levels as nominal categories rather than ordered stages of disease progression, leading to clinically inconsistent predictions. Third, although XAI techniques such as Grad-CAM have been proposed, their integration into DFU severity assessment remains limited, and visual explanations are rarely aligned with or validated against clinical reasoning. Finally, relatively few studies assess model latency, scalability, and robustness in real-world telemedicine settings, where efficiency, transparency, and reliability are equally essential.

Addressing these limitations requires a unified approach that moves beyond isolated improvements in individual components. In particular, the absence of frameworks that jointly integrate precise lesion segmentation, ordinal-aware severity grading, and explainable decision-making represents a key gap in current DFU assessment research.

This study addresses this gap by proposing an explainable hybrid deep learning framework that integrates U-Net-based segmentation, ordinal-aware EfficientNet-B3 classification, and Grad-CAM visualization within a single pipeline [23]. The proposed model is explicitly designed to achieve high accuracy, interpretability, and computational efficiency, enabling trustworthy and real-time deployment in telemedicine and cloud-based healthcare systems.

F.COMPARISON WITH ALTERNATIVE HYBRID DEEP LEARNING FRAMEWORKS

Several hybrid deep learning frameworks have been proposed for DFU analysis, typically combining a segmentation backbone with a classification network to improve grading accuracy. Common configurations include U-Net paired with ResNet, DenseNet, or VGG-based classifiers, as well as attention-enhanced segmentation combined with conventional categorical classifiers. While these approaches report improvements over single-stage models, their comparative advantages and limitations warrant critical examination.

Compared with U-Net + ResNet hybrids, which emphasize deep residual feature extraction, the proposed U-Net + EfficientNet-B3 framework achieves comparable or superior classification accuracy with reduced computational complexity due to EfficientNet’s compound scaling strategy. DenseNet-based hybrids benefit from feature reuse but often incur higher memory overhead, limiting deployment feasibility in telemedicine and mobile environments. VGG-based hybrids, although structurally simple, typically underperform in both accuracy and efficiency due to their large parameter count and lack of modern optimization mechanisms.

A key distinction of the proposed framework lies in its integration of ordinal-aware learning, which is absent from most existing hybrid DFU systems. Prior hybrid models generally treat severity grading as a nominal classification task, leading to clinically inconsistent errors between non-adjacent grades. By contrast, the use of consistent rank logits in the EfficientNet-B3 classifier enforces ordinal constraints aligned with the Meggitt–Wagner scale, reducing large-grade misclassifications that are common in alternative hybrids.

Furthermore, while some hybrid frameworks incorporate post hoc visualization, explainability is rarely treated as a core design objective. In contrast to existing hybrids that prioritize numerical accuracy alone, the proposed pipeline embeds Grad-CAM-based visual explanations directly into the classification stage, enabling verification of model attention and supporting clinician trust. This combination of segmentation-guided input, ordinal-aware grading, and explainable output differentiates the proposed approach from previously reported hybrid DFU frameworks.

Overall, the comparative analysis indicates that the U-Net + EfficientNet-B3 configuration offers a more balanced trade-off between accuracy, interpretability, and computational efficiency than alternative hybrid architectures, making it better suited for real-world telemedicine deployment.

A.DATASET

A total of 6,500 high-resolution DFU images were collected from several hospitals and outpatient centers under institutional ethical approval. Each image depicts plantar or dorsal views of diabetic feet under varying illumination, skin pigmentation, and ulcer morphology, ensuring a diverse representation of real-world conditions.

The dataset used in this study consists of 6,500 clinical images of DFUs collected from multiple hospitals and outpatient wound-care centers through institutional collaboration. All data were obtained under approved institutional review procedures, with patient identifiers removed prior to analysis to ensure anonymity and compliance with ethical standards for secondary clinical data use. The study did not involve direct patient interaction, and informed consent was waived in accordance with institutional and national research ethics guidelines. Due to patient privacy and data-sharing restrictions imposed by participating institutions, the dataset is not publicly available; however, it may be accessed for academic research purposes upon reasonable request and subject to institutional approval.

Each sample includes two forms of annotation:

• •Segmentation masks: binary images labeling ulcer pixels as 1 and background pixels as 0. Masks were manually drawn by trained medical annotators and verified by two wound-care specialists. Manual annotation remains the clinical gold standard for medical image segmentation.
• •Severity grades: ordinal labels from Grade 0 to 5 following the Meggitt–Wagner classification system, assigned by at least two certified clinicians. Discrepancies were resolved by consensus review to ensure consistency.

The dataset is divided using stratified sampling to preserve class balance: 70% for training, 15% for validation, and 15% for independent testing. To enhance generalization, on-the-fly data augmentation is performed only on the training set, including random horizontal/vertical flips, ±20° rotations, brightness/contrast jitter, and Gaussian noise addition. The original class distribution was maintained to avoid bias.

All experiments are conducted in Python 3.10 using TensorFlow 2.9 and Keras 2.9 on a workstation equipped with an NVIDIA RTX 3060 GPU (12 GB VRAM), Intel Core i7 CPU, and 32 GB RAM. To support reproducibility, the implementation details, training configurations, and model weights will be made available to the research community upon reasonable request, subject to institutional and ethical approval, with a public code repository planned for release following manuscript acceptance.

B.HYBRID DEEP LEARNING PIPELINE

The proposed framework (Fig. 1) integrates segmentation, classification, and interpretability into a unified deep learning pipeline for automated DFU severity assessment. The system follows four sequential stages:

• 1.Image input,
• 2.Wound segmentation via U-Net,
• 3.ROI extraction and severity classification using EfficientNet-B3 enhanced by Grad-CAM visualization.

Fig. 1. Overview of the proposed hybrid deep learning framework for diabetic foot ulcer (DFU) assessment.

Each component was designed to preserve spatial accuracy, semantic richness, and transparency while maintaining computational efficiency.

The framework implements a four-stage pipeline: (1) acquisition of a raw clinical foot image; (2) automated ulcer segmentation using a U-Net-based model to generate a binary wound mask; (3) ROI extraction guided by the segmentation output; and (4) severity classification using an EfficientNet-B3 network with ordinal-aware learning. Grad-CAM visualization is applied to highlight image regions that most influence the predicted wound grade (0–5), enabling visual verification of model reasoning and supporting transparent, clinically meaningful DFU assessment.

C.TRAINING AND OPTIMIZATION STRATEGY

Both modules are trained independently and then integrated for end-to-end evaluation. All models use the AdamW optimizer with default β-parameters (β₁ = 0.9, β₂ = 0.999) and decoupled weight decay = 1 × 10⁻⁴. The learning rate is reduced by a factor of 0.1 after 10 epochs of plateau. Early stopping was applied when validation loss did not improve for 10 consecutive epochs. Data augmentation strategy:

• •Segmentation: random rotations (±20°), flips, and color jitter.
• •Classification: random brightness/contrast shifts and CLAHE normalization.

These augmentations simulate real clinical variability (lighting, orientation, and camera noise) and improve robustness. Training duration: approximately 6 hours for segmentation and 5 hours for classification on the specified hardware. Model checkpoints are stored after each epoch for reproducibility.

D.EVALUATION METRICS

To ensure objective and comprehensive performance assessment, four groups of metrics were applied:

• 1.Segmentation performance–evaluated by
  • ○IOU:
    IoU=|P∩G||P∪G|
    Dice coefficient:
    Dice=2|P∩G||P|+|G|
    where P and G represent predicted and ground-truth masks, respectively. High scores indicate strong spatial overlap and boundary fidelity.
• 2.Classification performance–evaluated by
  • ○Accuracy, precision, recall, and macro-F1 score, which balance per-class performance on imbalanced medical datasets;
  • ○Confusion matrix to visualize inter-grade misclassifications and verify ordinal consistency.
• 3.Hybrid pipeline evaluation–classification accuracy is compared under two conditions: (a) with segmentation-based ROI input and (b) with unsegmented full images. The comparison quantifies the contribution of lesion localization to grading accuracy. Additionally, error propagation analysis examines how segmentation quality affects classification results, following methods outlined in prior hybrid studies.
• 4.Explainability evaluation qualitative analysis of Grad-CAM heatmaps assessed whether model attention corresponded to clinically relevant wound regions. Heatmaps showing focus on ulcer centers and edges were considered valid explanations, whereas diffuse or irrelevant activations indicated potential model bias.

All quantitative metrics were averaged over the test set with 95% confidence intervals computed via bootstrapping.

E.IMPLEMENTATION AND REPRODUCIBILITY

All experiments are executed in a controlled environment using the following configurations:

• •Programming language: Python 3.10
• •Libraries: TensorFlow 2.9, Keras 2.9, OpenCV 4.7, NumPy 1.26, Matplotlib 3.8
• •Hardware: NVIDIA RTX 3060 GPU (12 GB), Intel Core i7 CPU, 32 GB RAM
• •Operating system: Windows 11 64-bit

Random seeds are fixed across TensorFlow, NumPy, and Python to ensure deterministic outcomes. All training and evaluation scripts, along with preprocessing pipelines and model weights, can be provided by the authors for academic reproduction. No external code packages beyond the versions listed were used. Pre-registration was not required, as this work does not involve human or animal experimentation but utilizes anonymized secondary clinical images with institutional clearance.

A.SEGMENTATION PERFORMANCE

The segmentation module forms the basis of the hybrid architecture, enabling lesion-focused classification. Its precision determines how effectively the classifier isolates meaningful features from noisy image backgrounds.

Using IOU and Dice coefficient metrics, the baseline U-Net achieved an IOU of 0.843 and a Dice coefficient of 0.912, demonstrating strong alignment between predicted masks and expert annotations [25]. When the Attention U-Net variant was applied, these values increased to 0.861 (IOU) and 0.924 (Dice), showing measurable gains in boundary precision and robustness against texture variations.

These metrics are comparable with high-performing biomedical segmentation networks in other medical domains such as skin lesion detection (Dice ≈0.92) and diabetic retinopathy segmentation (Dice≈0.90) indicating the adaptability of U-Net-based architectures to diabetic foot imaging. The improvement from attention mechanisms can be attributed to the model’s ability to assign higher weights to informative pixels while suppressing irrelevant context, as illustrated in Fig. 4.

Fig. 4. Segmentation results: Ground truth vs. predicted masks.

Visual comparison between real DFU samples and corresponding predictions from U-Net and Attention U-Net. Each column displays a distinct ulcer case, showing how the attention mechanism refines wound boundary detection under challenging conditions such as lighting variation, overlapping toes, or partially occluded lesions.

Qualitative analysis shows that segmentation accuracy is most affected by ulcer appearance variability; bright reflections from moist tissue or darkened necrotic regions can confuse the model [26]. Nonetheless, the Attention U-Net demonstrated stable performance even under such circumstances, producing continuous and anatomically consistent contours.

This high segmentation accuracy ensures that the subsequent classification network receives accurate ROI inputs, which improves both computational efficiency and clinical interpretability. Accurate wound boundary localization also facilitates longitudinal monitoring, allowing automatic computation of wound area, perimeter, and healing rate over time metrics that are essential in diabetic care.

B.CLASSIFICATION PERFORMANCE

The classification module predicts DFU severity grades (0–5) according to the Meggitt–Wagner scale. The model achieved an overall accuracy of 92.3% and a macro-F1 score of 0.904, demonstrating high predictive performance across all severity levels.

When compared with baseline architectures, EfficientNet-B3 outperformed ResNet-50 (accuracy = 89.2%, F1 = 0.872) and VGG16 (accuracy = 86.7%, F1 = 0.841) [27]. These results confirm the effectiveness of compound scaling in EfficientNet, which optimizes network depth, width, and resolution simultaneously to improve feature extraction without excessive computational cost.

Performance gains are most pronounced in mid- to late-stage ulcers (Grades 3–5), where morphological heterogeneity is highest. These cases often involve complex textural cues such as necrotic patches, deep fissures, or infection indicators that require robust multi-scale feature extraction. The classifier also performed well for lower grades, distinguishing mild abrasions (Grade 1) from healthy tissue (Grade 0) with minimal false positives in Fig. 5.

Fig. 5. Confusion matrix illustrating the performance of the ordinal-aware EfficientNet-B3 classifier for diabetic foot ulcer (DFU) severity grading.

Rows represent the ground-truth severity grades (0–5), while columns indicate the predicted grades. The strong concentration of values along the diagonal indicates high classification accuracy, whereas the majority of misclassifications occur between adjacent grades (e.g., 2 ↔ 3 and 3 ↔ 4), demonstrating ordinal consistency and validating the effectiveness of the CORAL loss.

The ordinal-aware training proved crucial. Standard categorical cross-entropy loss treats severity classes as independent categories, allowing large jumps between unrelated grades. By contrast, CORAL loss enforces hierarchical consistency, aligning model behavior with the progressive nature of diabetic wound pathology [28]. This approach penalizes distant misclassifications more heavily, thus reducing clinically unacceptable grade reversals (1 → 5) in Fig. 6.

Fig. 6. ROC curves showing the classification performance of the proposed model across diabetic foot ulcer (DFU) severity grades.

ROC curves are presented for each grade, with area under the curve (AUC) values indicating balanced sensitivity and specificity (AUC > 0.90 for all classes). The results demonstrate robust discriminative capability across severity levels, with slightly reduced separability for intermediate grades due to overlapping clinical characteristics.

The high AUC values indicate that the model captures discriminative features effectively across varying ulcer morphologies. The ROC pattern also confirms that mid-grades (2–4) exhibit slightly lower separability than extreme grades (0 or 5), reflecting natural overlap in clinical presentation.

Comparatively, earlier studies using CNN-only classification reported accuracies of 82–88% for DFU grading without segmentation or ordinal learning. Our model’s improved performance underscores the importance of combining region-focused inputs and ordinal constraints.

C.END-TO-END PIPELINE EVALUATION

The hybrid system’s innovation lies in combining segmentation and classification in a single, explainable workflow. Evaluation focused on three aspects: holistic accuracy, module dependency, and error propagation.

D.MODEL INTERPRETABILITY AND VISUAL EXPLANATION

Explainability bridges AI output and clinical reasoning. Using Grad-CAM, visual attention maps were generated for all predictions to validate whether the network focused on medically relevant areas.

E.INFERENCE TIME AND DEPLOYMENT FEASIBILITY

In telemedicine applications, inference speed is a critical determinant of system usability and clinical practicality. On an NVIDIA RTX 3060 GPU, the proposed hybrid pipeline achieved an average inference time of 1.82 s per image, satisfying real-time clinical requirements. The computational cost of individual processing stages was 0.98 s for ulcer segmentation using U-Net and 0.84 s for severity classification using EfficientNet-B3, resulting in a combined end-to-end inference time of 1.82 s.

Figure 9 demonstrates the superior performance of our hybrid framework compared to traditional methods. According to Wang et al., while traditional algorithms like SVM perform well on small datasets, deep learning models (CNN) are more robust for large-scale image recognition. By transitioning from manual feature engineering to automated feature learning through deep hidden layers, our system more accurately captures complex wound morphologies, ensuring higher precision in diabetic ulcer grading.

Compared with traditional machine learning methods, including Random Forest (3.45 s) and Support Vector Machine (4.02 s), the hybrid framework provides a substantially improved balance between accuracy and computational efficiency. A detailed comparison of inference times across different models is presented in Table I. Although the hybrid approach is marginally slower than a single-stage CNN (0.83 s), the additional inference time is justified by the model’s enhanced interpretability, ordinal-aware severity grading, and overall diagnostic reliability.

Table I. Inference time comparison across different models

Average per-image inference times for different architectures. The hybrid U-Net + EfficientNet-B3 pipeline achieves sub-2-second latency while outperforming traditional baselines.

The design’s modularity also allows optimization for deployment on mobile and edge devices, where hardware constraints are stricter. By pruning non-critical layers and applying quantization, latency could be further reduced to under 1.5 s without notable accuracy loss.

F.COMPARISON WITH EXISTING HYBRID AND ENSEMBLE DFU APPROACHES

Recent studies have explored hybrid and ensemble-based deep learning strategies for DFU analysis in an effort to overcome the limitations of single-stage models. Typical hybrid approaches combine a segmentation network, most commonly U-Net or its variants, with a downstream classifier such as ResNet, DenseNet, or VGG to improve lesion-focused grading. Ensemble methods further aggregate predictions from multiple classifiers or feature extractors to enhance robustness and reduce variance. While these strategies often report incremental performance gains, their comparative evaluation remains limited in scope and clinical relevance.

Many existing hybrid DFU frameworks primarily benchmark against single CNN baselines without systematic comparison to alternative hybrid or ensemble configurations. As a result, improvements are often attributed to architectural complexity rather than principled design choices. In addition, ensemble methods, although capable of boosting accuracy, introduce increased computational cost, higher memory requirements, and reduced interpretability, which constrain their suitability for real-time telemedicine deployment. These trade-offs are rarely quantified explicitly in prior work, particularly with respect to inference latency and explainability.

In contrast to most reported hybrid and ensemble approaches, the proposed U-Net + EfficientNet-B3 framework emphasizes a balanced integration of segmentation-guided feature extraction, computational efficiency, and clinical interpretability. EfficientNet-based classifiers offer competitive performance with substantially fewer parameters than traditional deep CNNs, enabling faster inference compared with ensemble pipelines that rely on multiple backbone networks. Furthermore, while ensemble DFU models typically provide only aggregated predictions, the proposed framework retains transparency through Grad-CAM visualization, allowing clinicians to inspect the basis of individual severity assessments.

Another key limitation of existing hybrid and ensemble DFU systems is the lack of ordinal-aware severity modeling. Most approaches treat wound grading as a nominal classification task, even when using advanced architectures or ensembles. This limitation persists across both hybrid and ensemble designs and can result in clinically inconsistent errors between non-adjacent severity levels. By incorporating ordinal-aware learning within the classification stage, the proposed framework addresses a gap that remains largely unexplored in prior DFU hybrid and ensemble research.

Overall, although hybrid and ensemble methods have demonstrated the potential of multi-stage DFU analysis, existing approaches often prioritize accuracy improvements without adequately considering interpretability, ordinal consistency, or deployment feasibility. The proposed framework advances beyond these limitations by offering a unified, explainable, and computationally efficient hybrid design that is better aligned with clinical workflows and telemedicine requirements.

G.ARCHITECTURAL COMPLEXITY AND COMPUTATIONAL COST

Beyond inference latency, architectural complexity is a key consideration for practical deployment. The proposed U-Net + EfficientNet-B3 framework was designed to balance representational capacity with computational efficiency. EfficientNet-B3 achieves high classification performance with fewer parameters than conventional ResNet or DenseNet backbones, resulting in reduced memory usage and faster inference compared to deeper hybrid architectures. In contrast, ensemble-based DFU models require multiple forward passes and increased storage, significantly increasing time and space complexity. By avoiding ensemble aggregation and leveraging parameter-efficient backbones, the proposed framework maintains sub-2-second end-to-end inference while supporting explainability and ordinal-aware grading, making it more suitable for deployment in telemedicine and resource-constrained environments.

H.DISCUSSION AND CLINICAL IMPLICATIONS

While the quantitative results demonstrate strong segmentation accuracy, classification performance, and computational efficiency, the broader significance of the proposed framework lies in its alignment with clinical workflows and its potential to support trustworthy AI-assisted decision-making in real-world telemedicine settings. Rather than functioning as a purely predictive model, the hybrid pipeline is designed to emulate the sequential reasoning process used by wound-care specialists, thereby enhancing interpretability, clinical relevance, and adoption potential.

I.CLINICAL VALIDITY AND DIAGNOSTIC WORKFLOW ALIGNMENT

The proposed hybrid framework mirrors standard clinical practice by decomposing DFU assessment into discrete yet interdependent stages. Clinicians typically begin by visually identifying and localizing the ulcer, followed by focused examination of the lesion region to assess depth, tissue condition, and signs of infection before assigning a severity grade. By first performing ulcer segmentation, then restricting analysis to the extracted ROI, and finally applying ordinal-aware severity grading, the model follows this same diagnostic sequence. This structured workflow improves predictive reliability and provides interpretable intermediate outputs such as segmentation masks and attention maps that clinicians can review to support verification and informed decision-making [34].

J.DIAGNOSTIC CONSISTENCY AND REDUCED VARIABILITY

Clinical wound grading is known to suffer from inter-rater variability due to subjective interpretation of lesion severity. By providing quantitative, image-based severity grading, the proposed framework functions as a decision-support tool that harmonizes assessments across practitioners [35]. This consistency supports the standardization of DFU evaluation and reduces inter-observer variability, which has been identified as a key requirement for scalable and reliable deployment of AI systems in clinical practice [36]. Improved standardization may enhance patient outcomes by enabling timely escalation of severe cases and more consistent monitoring of wound progression.

K.INTEGRATION INTO TELEMEDICINE PLATFORMS

Given its real-time inference capability (under 2 s per image) and modular design, the framework is well suited for integration into telemedicine platforms. Patients can capture wound images using smartphones and receive immediate severity assessments, which clinicians can verify through the accompanying segmentation and Grad-CAM visualizations. Such AI-assisted remote monitoring enables early intervention and continuous care, particularly in rural or low-resource settings where access to wound-care specialists is limited. In this context, standardized and XAI systems are especially valuable, as reliability, transparency, and reproducibility are essential for safe and effective remote clinical decision-making.

L.SCALABILITY AND FUTURE ADAPTATIONS

The modular structure of the proposed pipeline facilitates scalability and domain adaptation. The segmentation component can be retrained for other chronic wound types, such as pressure ulcers or venous leg ulcers, while the classification module can be extended to predict infection status or healing outcomes. Future iterations may also integrate multimodal data, including clinical records, biomarkers, and longitudinal wound images, to enable more comprehensive and patient-centered decision support. These characteristics position the framework as a flexible foundation for broader deployment in intelligent telemedicine and digital health ecosystems.

M.LIMITATIONS AND FUTURE WORK

Despite promising results, several limitations must be acknowledged.

• 1.Dataset Scope: The dataset, while large, was collected under semi-controlled conditions. Real-world variability lighting differences, occlusions, or camera quality may affect generalizability.
• 2.Grad-CAM Resolution: Grad-CAM provides coarse heatmaps that highlight attention areas but not precise boundaries. Incorporating finer-grained methods like Score-CAM or Layer-wise Relevance Propagation could enhance interpretability.
• 3.Quantitative Explainability Validation: Although Grad-CAM visualizations indicate alignment between model attention and clinically relevant ulcer regions, the explainability analysis in this study remains primarily qualitative. No quantitative measure was used to assess the agreement between model-derived attention and expert clinical reasoning. Future work will address this limitation by computing overlap metrics, such as IoU, between Grad-CAM activation maps and clinician-annotated regions of interest to enable objective validation of explanation fidelity.
• 4.Error Propagation: As shown in Section IV.C.3, segmentation errors can degrade classification accuracy. Future work should explore joint end-to-end training to minimize error transfer between modules.
• 5.Limited Clinical Validation: Although model performance is quantitatively strong, prospective clinical trials are required to confirm effectiveness in live telemedicine workflows.

Future research directions include federated learning for privacy-preserving model training, active learning for annotation efficiency, and integration of temporal data for wound healing prediction. The hybrid deep learning framework successfully combines segmentation, ordinal classification, and interpretability for DFU assessment. Its main achievements include:

• •High accuracy: IOU = 0.861, Dice = 0.924, accuracy = 92.3%, macro-F1 = 0.904.
• •Explainability: Grad-CAM visualization aligns AI focus with clinical reasoning.
• •Efficiency: Sub-2-second inference time suitable for telemedicine.
• •Clinical utility: Modular, interpretable design mirrors human diagnostic workflow.

This study demonstrates that integrating technical innovation with clinical insight produces AI systems that are not only accurate but also transparent, efficient, and trustworthy. The results lay the groundwork for the next generation of intelligent telemedicine tools capable of transforming diabetic wound management.