Class	Japanese	English
ADM	後天性真皮メラノサイトーシス	Acquired Dermal Melanocytosis
BCC	基底細胞癌	Basal Cell Carcinoma
Ephelis	雀卵斑	Freckles
Melanoma	悪性黒色腫	Malignant Melanoma
Melasma	肝斑	Melasma
Nevus	色素性母斑	Melanocytic Nevus
SK	脂漏性角化症	Seborrheic Keratosis
Solar Lentigo	日光黒子	Solar Lentigo

Available Model Configurations

このリポジトリには2つのモデル構成が含まれています：

Configuration	Models	Folder	Use Case
Base Models	20 models	`base_models/`	シンプルなアンサンブル、軽量版
Full Ensemble	35 models	Root folders	最高精度、バイナリ精緻化あり

Base Models (20 models) - バイナリなし版

base_models/
├── efficientnet_b4_ham10k/      # EfficientNet-B4 × 5
├── efficientnet_b7_imagenet/    # EfficientNet-B7 × 5
└── transformers/                 # ViT × 5, Swin × 5

Full Ensemble (35 models) - フル版

├── efficientnet_b4_ham10k/      # EfficientNet-B4 × 5
├── efficientnet_b7_imagenet/    # EfficientNet-B7 × 5
├── transformers/                 # ViT × 5, Swin × 5
└── binary_classifiers/          # Binary classifiers × 15

Performance

Metric	Value
Test Accuracy	95.92%
Macro Precision	96.09%
Macro Recall	95.56%
Macro F1-Score	95.77%
Melanoma Sensitivity	100%
Melanoma Specificity	100%

Per-Class Performance

Class	Precision	Recall	F1-Score
ADM	96.30%	96.30%	96.30%
Basal Cell Carcinoma	87.50%	91.30%	89.36%
Ephelis	100.00%	94.44%	97.14%
Malignant Melanoma	95.24%	100.00%	97.56%
Melasma	97.14%	100.00%	98.55%
Nevus	100.00%	94.44%	97.14%
Seborrheic Keratosis	95.65%	88.00%	91.67%
Solar Lentigo	96.88%	100.00%	98.41%

Model Architecture

7 Model Types (35 total models)

Model	Pre-training	Parameters	Count
EfficientNet-B4	ImageNet → HAM10000	~19M	× 5 folds
EfficientNet-B7	ImageNet	~66M	× 5 folds
Vision Transformer (ViT)	ImageNet	~86M	× 5 folds
Swin Transformer	ImageNet	~88M	× 5 folds
Binary: ADM vs Nevus	-	~19M	× 5 folds
Binary: Nevus vs Melanoma	-	~19M	× 5 folds
Binary: Solar vs Ephelis	-	~19M	× 5 folds

Ensemble Strategy

Stage 1: Base Ensemble (20 models)

Average softmax probabilities from all base models

Stage 2: Binary Refinement (15 models) - Optional

Apply binary classifiers to refine confusing pairs:
- ADM ↔ Nevus
- Nevus ↔ Melanoma
- Solar Lentigo ↔ Ephelis

Usage

Installation

pip install torch torchvision timm pillow huggingface_hub

Download Models

python

from huggingface_hub import snapshot_download

# Option 1: Download base models only (20 models, ~5GB)
snapshot_download(
    repo_id="kindai-derma-ai/skin-lesion-ensemble",
    local_dir="./models",
    allow_patterns="base_models/**"
)

# Option 2: Download full ensemble (35 models, ~6.4GB)
snapshot_download(
    repo_id="kindai-derma-ai/skin-lesion-ensemble",
    local_dir="./models",
    ignore_patterns="base_models/**"  # Avoid duplicates
)

# Option 3: Download everything
snapshot_download(
    repo_id="kindai-derma-ai/skin-lesion-ensemble",
    local_dir="./models"
)

Inference with Base Models

python

import torch
import timm
from PIL import Image
from torchvision import transforms

# Load model (example with EfficientNet-B4)
model = timm.create_model('efficientnet_b4', pretrained=False, num_classes=8)
checkpoint = torch.load('models/base_models/efficientnet_b4_ham10k/efficientnet_b4_ham10k_fold0.pth')
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

# Preprocessing
transform = transforms.Compose([
    transforms.Resize((384, 384)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

# Predict
image = Image.open('skin_lesion.jpg').convert('RGB')
input_tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    output = model(input_tensor)
    probabilities = torch.softmax(output, dim=1)
    predicted_class = torch.argmax(probabilities, dim=1).item()

class_names = ['ADM', 'BCC', 'Ephelis', 'Melanoma', 'Melasma', 'Nevus', 'SK', 'Solar_Lentigo']
print(f"Prediction: {class_names[predicted_class]}")
print(f"Confidence: {probabilities[0][predicted_class]:.2%}")

File Structure

models/
├── base_models/                      # Base models only (20 models)
│   ├── efficientnet_b4_ham10k/
│   ├── efficientnet_b7_imagenet/
│   └── transformers/
│
├── efficientnet_b4_ham10k/           # Full ensemble components
├── efficientnet_b7_imagenet/
├── transformers/
└── binary_classifiers/               # Binary refinement (full ensemble only)

Training Details

Input Size: 384 × 384
Loss Function: Focal Loss (α=1, γ=2)
Optimizer: AdamW (lr=1e-4, weight_decay=0.01)
Scheduler: CosineAnnealingLR
Batch Size: 16
Epochs: 50 (with early stopping, patience=15)
Cross-Validation: 5-fold stratified

Data Augmentation

RandomHorizontalFlip
RandomVerticalFlip
RandomRotation (±20°)
ColorJitter (brightness, contrast, saturation, hue)
RandomAffine (translation, scale)

Limitations and Intended Use

Intended Use

Research and educational purposes
Dermatology AI research
Clinical decision support (with physician oversight)

Limitations

Not for clinical diagnosis: This model is intended as a decision support tool and should not replace professional medical diagnosis
Dataset bias: Trained primarily on Japanese patient data; performance may vary on other populations
Image quality dependency: Performance depends on image quality and standardized acquisition conditions

Out-of-Scope Use

Autonomous clinical diagnosis without physician review
Use on non-dermatoscopic images
Commercial deployment without proper validation

Citation

bibtex

@misc{skin_lesion_ensemble_2024,
  title={7-Model Ensemble for 8-Class Skin Lesion Classification},
  author={Kindai Dermatology AI Lab},
  year={2024},
  howpublished={\url{https://huggingface.co/kindai-derma-ai/skin-lesion-ensemble}}
}

License

This model is released under CC BY-NC 4.0.

Academic/Research Use: Freely available
Commercial Use: Contact authors for licensing

Acknowledgments

Created: December 2024 Organization: Kindai Dermatology AI Lab