**Background:** The emergence of 6G-enabled Internet of Medical Things (IoMT) offers opportunities for real-time medical diagnosis, but challenges remain in processing large volumes of medical CT images efficiently. Deep learning models can extract features from medical images, but the high dimensionality of extracted features can degrade classification performance and increase computational burden. Metaheuristic optimization algorithms can address this by selecting only the most relevant features, though existing methods suffer from premature convergence in large search spaces.
**Methods:** The proposed framework consists of two phases. First, a pretrained MobileNetV3-Large model (approximately 3 million trainable parameters) was fine-tuned on each medical imaging dataset for 100 epochs with batch size 32, using RMSprop optimizer (learning rate 1×10⁻⁴) and early stopping (20 epochs). The model extracted 128-dimensional feature vectors from each image via 1×1 pointwise convolution layers. Second, a novel hybrid feature selection algorithm (AOAHG) was developed by integrating the arithmetic optimization algorithm (AOA) with hunger games search (HGS) operators. The AOAHG algorithm uses a probability-based switching mechanism: when the probability Pi exceeds the Math Optimizer Accelerated (MOA) value, AOA operators (division/multiplication for exploration) are applied; otherwise, HGS operators (using hunger weights W1 and W2) are applied for enhanced exploitation. The fitness function balances classification accuracy (using SVM validation loss) and feature reduction ratio, controlled by parameter λ. The algorithm was evaluated with population size 50 and 20 iterations across all experiments. Four medical imaging datasets were used: ISIC-2016 (1179 skin lesion images: 248 malignant, 1031 benign; 70/30 train/test split), PH2 (200 dermoscopy images: 80 common nevus, 80 atypical nevus, 40 melanoma; 85/15 split), WBC (12,444 microscopic blood cell images: 3120 eosinophils, 3103 lymphocytes, 3098 monocytes, 3123 neutrophils; 80/20 split), and OCT (84,484 retinal OCT B-scans from 4686 patients: 8866 DME, 37,455 CNV, 11,598 drusen, 26,565 normal; 83,516 training, 968 testing).
**Key Results:** The AOAHG algorithm with SVM classifier achieved the best performance across all datasets. On ISIC-2016: accuracy 87.34%, precision 86.53%, recall 86.28%, F1-score 86.47%. On PH2: accuracy 96.43%, precision 96.44%, recall 96.43%, F1-score 96.43%, balanced accuracy 97.02%. On WBC: accuracy 88.62%, recall 88.62%, F1-score 88.80%, balanced accuracy 88.62%. On OCT: accuracy 99.69%, precision 99.69%, recall 99.69%, F1-score 99.69%, balanced accuracy 99.69%. Compared to nine other metaheuristic algorithms, AOAHG consistently ranked first or near-first across all metrics. The SVM classifier outperformed KNN, RF, and XGB classifiers overall, achieving an average accuracy of 92.48% across all datasets compared to 91.92% (KNN), 91.35% (XGB), and 91.34% (RF). The AOAHG algorithm also demonstrated competitive computational efficiency with an average execution time of 1.7031 seconds using SVM, faster than ArchOA (1.9405s), AO (2.3397s), HGS (2.3704s), MFO (2.4415s), CGO (2.7837s), GWO (3.1439s), BAT (3.3205s), AOA (3.3388s), and PSO (4.4469s).
**Clinical Implications:** The proposed 6G-enabled IoMT framework demonstrates that combining deep learning feature extraction (MobileNetV3) with hybrid metaheuristic feature selection (AOAHG) can achieve high-accuracy medical image classification suitable for real-time diagnostic applications. The lightweight design of MobileNetV3 (3 million parameters) makes it deployable on resource-constrained edge devices within 6G IoT infrastructure. The framework's ability to reduce feature dimensionality while maintaining or improving classification accuracy addresses the curse of dimensionality problem in medical imaging. The high performance on OCT (99.69% accuracy) and PH2 (96.43% accuracy) datasets suggests particular promise for retinal disease and skin cancer screening applications. However, the authors acknowledge that the framework is time- and memory-intensive, and future work should focus on simplification, investigating additional augmentation techniques, hyperparameter optimization of deep learning models, and combining multiple classification techniques.