A Lightweight Hybrid MobileNetV3-Random Forest Model for Diabetic Foot Ulcer Detection for Resource – Constrained Nigerian Clinics

Abstract

In Nigeria, Diabetic Foot Ulcer (DFU) is a severe consequence of diabetes that causes high rates of lower-leg amputations and death. Many primary care clinics throughout the nation depend on subjective, non-quantitative clinical judgments, which frequently lead in delayed diagnosis. The lightweight hybrid MobileNetV3-Random Forest model created in this study for automated DFU detection from smartphone-captured pictures, optimized for inexpensive devices found in resource-limited Nigerian clinics, confronts this problem. From several Nigerian medical sources, 1,721 anonymized DFU photographs including a variety of skin tones and ulcer grades were compiled. To increase robustness against lighting and orientation variances, images were normalized, resized to 224x224 pixels, and augmented. Reducing the image dimensionality from 28,224 to 96 principal components while retaining 95% of the variance, PCA was then utilized. The feature extractor, MobileNetV3 (pre-trained through transfer learning), generated a high-dimensional feature vector. For the ultimate DFU detection job, these condensed elements were input into a Random Forest classifier. With five-fold cross-validation employed to guarantee robustness, the dataset was split into 70/20/10 for training, testing and cross-validation respectively. Achieving 92.8% accuracy, 0.93 precision, 0.93 recall, 0.93 F1-score, and 0.9959 ROC-AUC, the hybrid model outperformed several heavier benchmarks in efficiency. Ablation studies validated that PCA decreased memory use by 31% on low-end devices and reduced inference time by about 47% (from 352 ms to 187 ms per image), with very little influence on predictive performance. This study offers a scalable, computationally economical, and understandable artificial intelligence framework supporting hybrid intelligence ideas that advances fair early diagnosis of diabetic foot ulcers in resource-poor medical environments.

DOI: https://doi.org/10.5281/zenodo.19543172