TY  - JOUR
AU  - Wang, Dingguo
AU  - Chen, Yudi
PY  - 2026
DA  - 2026/06/03
TI  - Comparative Study of Transfer Learning Strategies for Multi-Class Skin Lesion Classification: Architectures, Fine-Tuning, and Data Augmentation
JO  - ICCK Journal of Image Analysis and Processing
T2  - ICCK Journal of Image Analysis and Processing
JF  - ICCK Journal of Image Analysis and Processing
VL  - 2
IS  - 3
SP  - 153
EP  - 167
DO  - 10.62762/JIAP.2026.390206
UR  - https://www.icck.org/article/abs/JIAP.2026.390206
KW  - transfer learning
KW  - skin lesion classification
KW  - convolutional neural networks
KW  - fine-tuning strategies
KW  - data augmentation
KW  - HAM10000
KW  - medical image analysis
KW  - ablation study
AB  - Skin lesion classification is critical in dermatological diagnosis, where early and accurate identification of malignant lesions can significantly improve patient outcomes. Deep learning approaches, particularly transfer learning with pre-trained CNNs, have demonstrated remarkable performance in automated dermoscopic image analysis. However, the optimal configuration of transfer learning components---including backbone architecture, fine-tuning strategy, and data augmentation intensity---remains an open question. In this paper, we present a systematic comparative study on the HAM10000 dataset, evaluating three CNN architectures (ResNet50, DenseNet121, EfficientNet-B0), three fine-tuning strategies (full, partial, classifier-only), and three data augmentation strategies (basic, moderate, aggressive). Our experiments reveal: (1) all three architectures achieve comparable per-class F1-scores under basic augmentation, with no statistically significant differences (Welch's t-test, p > 0.05), despite EfficientNet-B0 reaching 100% overall validation accuracy; (2) full fine-tuning yields the highest accuracy (86.19%) and AUC (99.42%) at increased computational cost; (3) basic augmentation achieves the best performance (accuracy=91.43%, AUC=99.30%), while aggressive augmentation degrades results due to excessive distortion of medical image features. Ablation studies further demonstrate: label smoothing and inverse class frequency weights produce a small positive interaction effect; gradient clipping at norm 1.0 is essential for training stability (without it, training collapses); and a backbone learning rate of $5 \times 10^{-4}$ yields optimal partial fine-tuning performance. McNemar's test confirms no significant difference in BKL vs.\ MEL misclassification patterns between ResNet50 and EfficientNet-B0 ($p > 0.05$). These findings provide practical guidelines for configuring transfer learning pipelines in medical image classification.
SN  - 3068-6679
PB  - Institute of Central Computation and Knowledge
LA  - English
ER  - 

@article{Wang2026Comparativ,
  author = {Dingguo Wang and Yudi Chen},
  title = {Comparative Study of Transfer Learning Strategies for Multi-Class Skin Lesion Classification: Architectures, Fine-Tuning, and Data Augmentation},
  journal = {ICCK Journal of Image Analysis and Processing},
  year = {2026},
  volume = {2},
  number = {3},
  pages = {153-167},
  doi = {10.62762/JIAP.2026.390206},
  url = {https://www.icck.org/article/abs/JIAP.2026.390206},
  abstract = {Skin lesion classification is critical in dermatological diagnosis, where early and accurate identification of malignant lesions can significantly improve patient outcomes. Deep learning approaches, particularly transfer learning with pre-trained CNNs, have demonstrated remarkable performance in automated dermoscopic image analysis. However, the optimal configuration of transfer learning components---including backbone architecture, fine-tuning strategy, and data augmentation intensity---remains an open question. In this paper, we present a systematic comparative study on the HAM10000 dataset, evaluating three CNN architectures (ResNet50, DenseNet121, EfficientNet-B0), three fine-tuning strategies (full, partial, classifier-only), and three data augmentation strategies (basic, moderate, aggressive). Our experiments reveal: (1) all three architectures achieve comparable per-class F1-scores under basic augmentation, with no statistically significant differences (Welch's t-test, p > 0.05), despite EfficientNet-B0 reaching 100\% overall validation accuracy; (2) full fine-tuning yields the highest accuracy (86.19\%) and AUC (99.42\%) at increased computational cost; (3) basic augmentation achieves the best performance (accuracy=91.43\%, AUC=99.30\%), while aggressive augmentation degrades results due to excessive distortion of medical image features. Ablation studies further demonstrate: label smoothing and inverse class frequency weights produce a small positive interaction effect; gradient clipping at norm 1.0 is essential for training stability (without it, training collapses); and a backbone learning rate of \$5 \times 10^{-4}\$ yields optimal partial fine-tuning performance. McNemar's test confirms no significant difference in BKL vs.\ MEL misclassification patterns between ResNet50 and EfficientNet-B0 (\$p > 0.05\$). These findings provide practical guidelines for configuring transfer learning pipelines in medical image classification.},
  keywords = {transfer learning, skin lesion classification, convolutional neural networks, fine-tuning strategies, data augmentation, HAM10000, medical image analysis, ablation study},
  issn = {3068-6679},
  publisher = {Institute of Central Computation and Knowledge}
}