A retrospective technical evaluation has assessed a multi-stage machine learning framework designed to categorize wound severity from clinical photographs, combining deep-learning localization with a conventional classifier. The investigative team integrated the YOLOv3 object-detection network for region-of-interest localization, an AlexNet convolutional neural network architecture for deep feature extraction, and a support vector machine decision layer to classify wound regions into three distinct clinical severity tiers labeled green, yellow, and red. Across a verified held-out test set consisting of 150 regions of interest, the hybrid model attained an overall classification accuracy of 80.67 percent, demonstrating the technical feasibility of coupling specialized object detection with feature extraction and support vector machine classification on archival wound imagery.
Chronic and acute wounds represent an extensive clinical and operational burden across healthcare environments, frequently requiring ongoing specialist evaluation to guide therapeutic interventions. A cited 2018 retrospective analysis estimated that more than seven million Medicare beneficiaries experienced wounds, generating annual healthcare expenditures ranging between 30.1 billion and 100.8 billion dollars. The patient population evaluated in such clinical settings commonly presents with diverse wound presentations, including diabetic foot ulcers, venous leg ulcers, pressure injuries, and post-surgical wounds. In primary-care and geographically remote environments, timely access to dedicated wound specialists can be constrained, leading researchers to examine whether image-based computational tools could offer standardized documentation and objective triage support for clinical teams.
Image-based wound analysis encompasses several technically distinct objectives within the medical computing literature, such as detecting wound boundaries, identifying general wound etiologies, segmenting specific tissue compositions like granulation or slough, and categorizing tissue infection or ischemia. Prior studies had explored deep learning and transfer learning across various binary and multiclass tasks, including an earlier investigation by Anisuzzaman and colleagues that directly examined three-level severity classification on clinical photographs. The specific engineering gap addressed in the current investigation was not the initial creation of a wound-severity grading scheme, but the empirical evaluation of a unified, image-only pipeline that sequentially automates region-of-interest localization with YOLOv3, derives intermediate deep visual features, and substitutes the standard terminal softmax layer of a neural network with a support vector machine classifier.
The experimental dataset originated from the AZH Wound and Vascular Center situated in Milwaukee, Wisconsin, consisting of photographic wound records acquired using iPads and digital single-lens reflex cameras under unstandardized clinical lighting conditions. The archive incorporated clinical presentations of pressure ulcers, diabetic foot ulcers, and venous ulcers evaluated and labeled by an experienced wound care specialist. The three-category grading convention established at the center defined green as the least urgent clinical tier, yellow as requiring additional clinical attention, and red as denoting the most urgent severity status. The authors documented that the dataset derived from the same source image collection previously studied by Anisuzzaman and colleagues, noting aggregate archive counts ranging from 420 to 425 source images across historical records.
In the automated processing workflow, full-resolution clinical photographs were initially processed by a previously developed YOLOv3 localization model using a 95 percent detection confidence threshold to isolate bounding boxes surrounding suspicious wound lesions. Because larger or multi-focal wounds can yield multiple detected areas within a single patient photograph, individual images occasionally generated more than one distinct region of interest. The verified dataset split established an overall pool of 730 regions of interest, allocated into a development cohort of 580 regions of interest and an independent held-out test cohort of 150 regions of interest. Within the held-out test partition, the class breakdown comprised 62 red, 48 yellow, and 40 green wound regions.
To expand the training distribution while preventing synthetic patterns from biasing performance validation, data augmentation was applied strictly to the development partition after the primary split was conducted. Within the development subset, three peri-wound crop regions were created around each localized region using spatial padding boundaries of 50, 100, and 150 pixels, followed by horizontal flips, vertical flips, and rotations of 45 and 90 degrees. This procedure expanded the 580 development regions into 3,465 augmented samples, while the 150 held-out test samples remained unaugmented to maintain an untainted benchmark. The authors noted, however, that because source-image and patient identifiers were not fully documented in the surviving archive, the partition could not definitively guarantee patient-level isolation between sets.
Feature extraction was conducted using an AlexNet convolutional architecture chosen for its manageable implementation footprint, fixed 227 by 227 by 3 pixel input dimensionality, and established baseline utility in prior wound literature. The architecture processes cropped inputs through five sequential convolutional stages interspersed with pooling layers and fully connected layers. Instead of utilizing the network’s final softmax output layer to determine the severity classification, the researchers extracted the high-dimensional latent representation from the terminal deep layer and supplied these numerical feature vectors directly to a support vector machine classifier. Hyperparameters were selected on the development partition using a grid search guided by five-fold cross-validation, employing stochastic gradient descent with momentum, a batch size of 100, a learning rate of 0.001, and up to 100 training epochs.
Evaluated on the 150 held-out test regions of interest, the proposed hybrid framework correctly classified 121 samples, yielding a primary test accuracy of 80.67 percent alongside a macro-averaged precision of 80.75 percent, recall of 80.80 percent, and an F1-score of 80.90 percent. Additional test metrics on the held-out partition indicated a true-positive rate of 80.67 percent, a true-negative rate of 80.67 percent, and a Matthews correlation coefficient of 0.81. Across development folds, validation performance metrics ranged between 90 and 94 percent with a macro-average receiver operating characteristic area under the curve of 0.97, though the authors emphasized that development diagnostics should be distinguished from final held-out test results.
The study compared the hybrid model against four standard transfer-learning baseline architectures evaluated on the exact same held-out test split. In these benchmark tests, ResNet50 achieved an accuracy of 70.70 percent with a precision of 70.90 percent, recall of 70.60 percent, and an F1-score of 70.90 percent, representing the highest performance among standard comparators. VGG-16 recorded an accuracy of 65.50 percent, MobileNetV2 achieved 63.50 percent, and InceptionV3 reached 62.00 percent accuracy on the same test set. While the hybrid pipeline exhibited a descriptive 9.97 percentage point accuracy advantage over ResNet50, the authors emphasized that because granular per-image prediction lists were unavailable in the archive, formal paired statistical significance tests could not be completed.
The authors explicitly contextualized their findings as preliminary proof-of-concept evidence rather than confirmation of immediate clinical readiness. The investigation was retrospective and confined to technical image-classification metrics without measuring direct clinician-model agreement, diagnostic workflow integration, or patient outcomes. Significant methodological constraints included the relatively small single-center dataset, unstandardized photographic acquisition, lack of prospective validation, reliance on an older convolutional backbone, and missing software dependency and random-seed logs. Consequently, the researchers concluded that future evaluations must incorporate larger multi-center cohorts, verified patient-level separation, contemporary neural backbones, and prospective clinical trials before such automated systems can be deployed as decision-support aids.
Beyond the direct transfer-learning baselines evaluated on the held-out partition, earlier medical imaging publications examined alternative algorithmic structures on wound datasets. Across prior studies evaluating the AZH dataset, archived reports documented three-class accuracies of 79.90 percent for MobileNet, 78.80 percent for InceptionResNetV2, 78.40 percent for Quadratic support vector machines, 77.20 percent for InceptionV2, and 73.00 percent for ResNet101. On distinct diabetic foot ulcer archives, earlier models registered binary classification accuracies ranging between 70.00 percent and 93.00 percent for CNN and modified VGG-16 designs. The authors emphasized that because historical evaluations applied divergent task definitions, labeling conventions, and data partitions, such published figures serve solely as descriptive literature context rather than direct performance benchmarks against the unified localization and classification pipeline.
The investigation also detailed an unresolved data provenance discrepancy between historical records and the current analysis. Prior work by Anisuzzaman and colleagues reported a development and test split of 577 and 146 regions of interest, respectively, encompassing 154 green, 237 red, and 186 yellow samples. The present archived records listed 580 development and 150 test regions, comprising 155 green, 238 red, and 187 yellow samples. Because every category total differed by merely one to three regions, the researchers concluded that this alignment provides strong circumstantial evidence of substantial data overlap rather than an independently sampled cohort. Furthermore, archival clarification confirmed that exactly 121 out of the 150 held-out regions were correctly classified, establishing the verified 80.67 percent accuracy and refuting an unverified 97.33 percent calculation that arose from historical notation errors.
Technical execution was conducted in Python 3.9 utilizing the Keras framework on an Intel Core i9 workstation equipped with an NVIDIA 5070 graphics processing unit. The optimization protocol incorporated combined ReLU and Leaky ReLU activation functions, an epsilon value of 0.002, and 100 iterations per epoch. However, because specific momentum coefficients, dropout settings, random seeds, and software dependency versions were not retained in the archived files, future investigations must recover these experimental parameters directly from original execution scripts to achieve complete independent reproducibility.
Subject of Research: Automated image-based wound-severity classification using a hybrid deep learning and support vector machine framework
Article Title: A hybrid CNN and SVM framework for AI-driven clinical diagnosis and wound severity assessment
Article References: Alkanhel, R. I., Alshahrani, A., Muthanna, M., Ali, E., & Ahmed, M. M. (2026). A hybrid CNN and SVM framework for AI-driven clinical diagnosis and wound severity assessment. Frontiers in Medicine, 13, Article 1910774. https://doi.org/10.3389/fmed.2026.1910774
Image Credits: AI Generated
DOI: 10.3389/fmed.2026.1910774
Keywords: wound severity classification, deep learning, support vector machine, YOLOv3, AlexNet, medical imaging, clinical decision support, computer vision, wound care, transfer learning, retrospective study, AZH Wound and Vascular Center, image segmentation, convolutional neural network, classification accuracy, feature extraction
Cite Scienmag News
Blake Davidson. (November 17, 2026). Hybrid Deep Learning and SVM Framework Evaluates Image-Based Wound Severity. Scienmag. https://scienmag.com/hybrid-deep-learning-and-svm-framework-evaluates-image-based-wound-severity/
Blake Davidson. "Hybrid Deep Learning and SVM Framework Evaluates Image-Based Wound Severity." Scienmag, 17 November 2026, https://scienmag.com/hybrid-deep-learning-and-svm-framework-evaluates-image-based-wound-severity/. Accessed 17 November 2026.
Blake Davidson. "Hybrid Deep Learning and SVM Framework Evaluates Image-Based Wound Severity." Scienmag. November 17, 2026. https://scienmag.com/hybrid-deep-learning-and-svm-framework-evaluates-image-based-wound-severity/








