Comparative analysis of convolutional neural network architectures: the efficacy of individual learning strategies and hyperparameter tuning

A significant gap exists between the theoretical performance of convolutional neural network (CNN) architectures and their real-world efficacy in dermatoscopic skin disease diagnosis. The primary issue is the use of universal training pipelines that disregard the internal topology of specific models, leading to suboptimal accuracy and a high risk of false-negative melanoma predictions. The goal of this work is to conduct a comparative analysis of personalized learning strategies to enhance CNN efficiency for automated diagnostics. Experimental results demonstrate that the universal approach is inferior to personalized strategies. The maximum gain in the F1-metric is achieved by adapting hyperparameters to each architecture: for EfficientNet-B1, MixUp/CutMix and Automatic Mixed Precision (AMP) were applied to improve generalization ability; DenseNet-201 was trained using the OneCycleLR policy and Random Erasing augmentation for noise robustness; Inception-ResNet-V2 utilized an increased input resolution and the CosineAnnealingLR scheduler. On a stratified subset of the ISIC-2020 dataset, a statistically significant growth in the F1-score was recorded. A key achievement was a 60.8% reduction in false-negative melanoma predictions, while the proposed methods improved precision without a substantial increase in inference time or hardware requirements. The practical significance lies in confirming that the maximum quality improvement in diagnostics is achieved through architecture-specific tuning. The results show that implementing such personalized optimization strategies is expedient for enhancing the accuracy of medical diagnostic systems.