Manuscript received September 30, 2025; revised November 3, 2025; accepted January 5, 2026; published July 16, 2026.

Abstract—Automated skin lesion classification systems face significant challenges in handling severe class imbalance characteristic of clinical dermatological datasets. This study presents a comprehensive comparative analysis of two state-of-the-art convolutional neural network architectures—EfficientNet-B0 and MobileNetV3-Large—for multiclass skin lesion classification on the HAM10000 benchmark dataset. The investigation systematically addresses an extreme class imbalance ratio of 58.30:1 through the strategic application of RandomOverSampler methodology combined with transfer learning from ImageNet pretrained weights. We implement identical training protocols for fair architectural comparison, including progressive dropout regularization (0.3→0.2→0.1) and conservative data augmentation strategies. Comprehensive ablation studies demonstrate the cumulative effectiveness of each pipeline component, with class balancing contributing 10.39% improvement and transfer learning providing substantial performance gains. Experimental validation reveals that EfficientNet-B0 achieves 98.86% classification accuracy with balanced precision, recall, and F1-Scores of 98.89%, 98.86%, and 98.85%, respectively, while MobileNetV3-Large delivers comparable performance at 98.88% accuracy. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis validates that learned features align with clinical diagnostic criteria, demonstrating attention to lesion boundaries, color variations, and texture patterns. Compared to existing state-of-the-art methods on HAM10000, our approach achieves superior performance with fewer parameters (5.3 M), representing a marked advancement for resource-constrained deployment scenarios. These findings establish a validated framework for managing severely imbalanced medical imaging datasets and provide empirical guidance for efficient architecture selection in dermatological computer-aided diagnosis systems.

Keywords—skin lesion classification, class imbalance mitigation, transfer learning, EfficientNet architecture, MobileNetV3 framework, HAM10000 dataset, medical image analysis, dermatological diagnosis, Gradient-weighted Class Activation Mapping (Grad-CAM) interpretability, deep learning comparison

Cite: A.M. H. Pardede, Solikhun, and Juni Ismail, "Comparative Analysis of EfficientNet-B0 and MobileNetV3-large Architectures for Imbalanced Multiclass Skin Lesion Classification," Journal of Image and Graphics, Vol. 14, No. 4, pp. 551-566, 2026.

Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).