adeIaide/skin-lesion-7class-efficientnet

ISIC 2018 EfficientNet-B0 Skin Lesion Classifier (7-Class)

ISIC 2018 Task 3/HAM10000 피부 병변 이미지를 7개 클래스로 분류하는 EfficientNet-B0 기반 PyTorch 모델입니다. 메타데이터는 사용하지 않으며 RGB 피부경 이미지 한 장만 입력으로 받습니다.

교육·연구용 모델입니다. 의료 진단, 치료 결정 또는 환자 선별에 사용하지 마세요.

Model details

항목	값
Backbone	EfficientNet-B0, ImageNet pretrained
Input	RGB, 320 x 320
Output	7 classes
Fine-tuning	Last 4 feature blocks and classifier
Trainable parameters	3,707,855
Optimizer	AdamW
Learning rate	0.0002 for backbone and classifier
Epochs	20; best checkpoint at epoch 17
Main metric	Macro-F1

Class order:

0 MEL   Melanoma
1 NV    Melanocytic nevus
2 BCC   Basal cell carcinoma
3 AKIEC Actinic keratosis / intraepithelial carcinoma
4 BKL   Benign keratosis-like lesion
5 DF    Dermatofibroma
6 VASC  Vascular lesion

Evaluation

ISIC 2018 training images were split by lesion_id so photographs of the same lesion could not appear in both train and development sets. The internal development set contains 2,026 images. This is not an external clinical validation set.

Metric	Score
Macro-F1	0.7641
Balanced accuracy	0.7958
Accuracy	0.8638
Macro AUROC	0.9343

Brightness robustness

The same development images were evaluated after deterministic brightness changes. A factor of 1.0 is the original image.

Brightness factor	0.50	0.70	0.85	1.00	1.15	1.30	1.50
Macro-F1	0.5113	0.6678	0.7476	0.7641	0.7182	0.6937	0.5617

The full measurements are provided in brightness_metrics.csv.

Files

• model.safetensors: recommended tensor-only model weights.
• isic2018_7class_efficientnet_b0_320.pt: original self-contained PyTorch checkpoint.
• config.json: architecture, preprocessing, and class mapping.
• inference_example.py: standalone single-image inference example.
• brightness_metrics.csv: robustness results for seven brightness levels.

Usage

pip install -r requirements.txt
python inference_example.py path/to/dermoscopic_image.jpg

The example downloads model.safetensors from this repository automatically. It prints the predicted class and probabilities for all seven classes.

Training data and license

The model was trained on the ISIC 2018 Task 3/HAM10000 training data. The dataset is distributed under CC BY-NC terms. Use of these weights is therefore limited to non-commercial research and education, with attribution to the data creators.

• ISIC 2018 Challenge data: https://challenge.isic-archive.com/data/
• Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset. Scientific Data 5, 180161 (2018). https://doi.org/10.1038/sdata.2018.161
• Codella et al. Skin Lesion Analysis Toward Melanoma Detection 2018. https://arxiv.org/abs/1902.03368

Limitations

• The model recognizes only the seven ISIC 2018 classes and cannot identify other diseases or normal skin.
• Performance was measured on an internal lesion-disjoint split, not at an independent hospital or in prospective clinical use.
• The dataset is imbalanced, with melanocytic nevi forming the majority class.
• Predictions are sensitive to large brightness changes, as shown above.
• Results may not generalize across cameras, clinics, skin tones, body sites, or non-dermoscopic photographs.
• A high-confidence prediction is not a diagnosis.

Source code

• https://github.com/GREENAAPL/skin_project
• https://github.com/adeIaide/skin_project