The DOI: https://doi.org/10.21203/rs.3.rs-9285260/v1

Skin Lesion Classification with Deep Learning

Classifying 7 types of skin lesions from dermoscopic images using the HAM10000 dataset. The project compares CNN-based and Transformer-based architectures, and combines the two best-performing models in an ensemble.

Dataset ــ https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/DBW86T,

HAM10000 — 10,015 dermoscopic images across 7 classes:

• Melanocytic Nevi
• Melanoma
• Benign Keratosis
• Basal Cell Carcinoma
• Actinic Keratoses
• Vascular Lesions
• Dermatofibroma

The dataset is split at the patient level using GroupShuffleSplit on lesion_id to prevent data leakage between splits (72% train / 13% val / 15% test).

Models

Model	Test Acc	Balanced Acc	Macro F1	AUC-ROC
ResNet-50 (Baseline)	0.7590	0.7500	0.6681	0.9548
EfficientNet-B4	0.8507	0.7083	0.7380	0.9712
ConvNeXt-Base	0.8474	0.7634	0.7510	0.9428
Swin Transformer-Base	0.8507	0.7966	0.7719	0.9717
ViT-B/16	0.8566	0.7720	0.7741	0.9629
Ensemble (EfficientNet-B4 + Swin)	0.8657	0.7998	0.7856	0.9811

Training Setup

All models were trained with the following configuration:

• Optimizer: AdamW with weight decay 1e-2
• Loss: CrossEntropyLoss with label smoothing 0.1
• Mixed precision training (fp16)
• Gradient clipping (max norm 1.0)
• WeightedRandomSampler to handle class imbalance
• Early stopping (patience 7-8 epochs)

CNN models (ResNet-50, EfficientNet-B4, ConvNeXt-Base) used CosineAnnealingWarmRestarts. Transformer models (Swin, ViT) used Linear Warmup followed by Cosine Decay, which proved essential for stable training.

Notebooks

EDA/
    Skin_Cancer_EDA.ipynb

Models/
    Fine_Tuning_ResNet50.ipynb
    Fine_Tuning_EfficientNet_B4.ipynb
    Fine_Tuning_ConvNeXt_Base.ipynb
    Fine_Tuning_Swin_Transformer.ipynb
    Fine_Tuning_ViT_B16.ipynb
    Ensemble_EfficientNet_Swin.ipynb

Ensemble

The ensemble combines EfficientNet-B4 and Swin Transformer-Base using weighted average of softmax probabilities. The optimal weights (EfficientNet 0.4, Swin 0.6) were found via grid search on the validation set. The diversity between a CNN and a Transformer architecture is the main reason the ensemble outperforms each model individually.

Requirements

torch
timm
albumentations
torchmetrics
scikit-learn
pandas
numpy
matplotlib
seaborn

References

• Tschandl et al., The HAM10000 dataset (2018)
• Liu et al., Swin Transformer (2021)
• Dosovitskiy et al., An Image is Worth 16x16 Words (2020)
• Tan & Le, EfficientNet (2019)
• Liu et al., A ConvNet for the 2020s (2022)