Diagnostic precision vs. computational efficiency: Benchmarking lightweight and deep CNNs for melanoma classification

Mahfoodh, Ali

doi:10.30772/qjes.2026.173263.2076

	Diagnostic precision vs. computational efficiency: Benchmarking lightweight and deep CNNs for melanoma classification
Al-Qadisiyah Journal for Engineering Sciences
Articles in Press, Accepted Manuscript, Available Online from 28 July 2026
Document Type: Research Paper
DOI: 10.30772/qjes.2026.173263.2076
Author
Ali Mahfoodh^*
Biomedical Engineering Department, Near East University, Nicosia, Mersin 10, Turkey.
Abstract
The most lethal form of skin cancer is malignant melanoma; however, it is highly curable when detected at an early stage. Despite the exceptional diagnostic capabilities of Deep Convolutional Neural Networks (CNNs), it is essential to select the appropriate architectural paradigm in order to achieve the highest clinical safety, the lowest computation cost, and the best predictive performance. This investigation comprehensively evaluates the binary classification of dermoscopic images using five distinct pre-trained CNNs (SqueezeNet, GoogleNet, ResNet-50, MobileNetV2, and EfficientNetB0) and a variety of structural paradigms. A balanced dataset of 9,605 images was utilized alongside transfer learning, stochastic spatial augmentations, and an optimized Adam-based training protocol. To achieve consistent predictive stability (Accuracy = 90.94% ± 0.13%), EfficientNetB0 was tested using a 5-Fold Cross-Validation protocol, and the rest of the networks were tested using hold-out method with minimal computational burden. These findings of the quantitative analysis have illustrated that the clinical compromises of the architectures were diverse. The Area Under the Curve (AUC) and Specificity (98.40%) values of ResNet-50 were the highest, suggesting that it effectively reduced false alarms. On the other hand, GoogleNet, by emphasizing patient safety, obtained the highest Sensitivity (90.80%). Additionally, the SqueezeNet micro-architecture was capable of preventing class-collapse as a result of the optimized training process, which resulted in an increase in accuracy to 91.40%. This was critically important. The findings of this study indicate that deep residual networks are the most appropriate for high-precision screening, while the deployment of lightweight networks of a suitable scale is highly viable and reliable in computational environments with limited resources.
Keywords
Cutaneous Melanoma; Convolutional Neural Networks (CNN); SqueezeNet; MobileNetV2; ResNet-50; GoogleNet

Statistics Article View: 71