Skin Lesion Classification (Benign vs Malignant)

⚠️ MEDICAL & REGULATORY DISCLAIMER

This model is for educational and research purposes only.

Must NOT be used in real-world/clinical applications (screening, triage, diagnosis, or treatment).

Not a medical device and not reviewed/approved by any regulator (e.g., FDA/MHRA).

Outputs must NOT replace clinical judgment. All suspicious lesions should be evaluated by a qualified clinician.

Model Description

This model classifies dermoscopic images of skin lesions as benign or malignant using a fine-tuned ResNet50 architecture.

Architecture: ResNet50 (pre-trained on ImageNet) + Custom Classification Head
Task: Binary Classification (Benign vs Malignant)
Dataset: HAM10000 (10,015 dermoscopic images)
Training: 25 epochs, Intel Core i5-12500H (CPU), ~4.5 hours
Input: 224×224×3 RGB images
Parameters: 24.7M total (15.6M trainable, 63.1%)

Performance Metrics (Test Set - 1,505 images)

Metric	Value
Accuracy	81.59%
ROC-AUC	0.8900
PR-AUC	0.6602
Precision	0.5121
Recall	0.8056
Loss	0.4426

Clinical Interpretation:

High Recall (80.56%): Catches 80% of malignant lesions (crucial for screening)
Moderate Precision (51.21%): Some false positives (acceptable for medical screening)
Strong ROC-AUC (0.89): Excellent discrimination ability

Usage

Installation

pip install tensorflow huggingface_hub pillow numpy

Complete Inference Script

# Import required libraries
from huggingface_hub import hf_hub_download
import tensorflow as tf
import numpy as np
from PIL import Image

# Download model from Hugging Face
print("Downloading model from Hugging Face...")
model_path = hf_hub_download(
    repo_id="devatreya/skin-lesion-resnet50",
    filename="resnet50_best.h5"
)

# Load model
print("Loading model...")
model = tf.keras.models.load_model(model_path)

# Preprocess image function
def preprocess_image(image_path):
    """Load and preprocess image for ResNet50"""
    img = Image.open(image_path).convert('RGB')
    img = img.resize((224, 224))
    img_array = np.array(img, dtype=np.float32)
    
    # ResNet50 preprocessing (ImageNet/Caffe-style)
    img_array = img_array[..., ::-1]  # RGB to BGR
    mean = [103.939, 116.779, 123.68]
    img_array[..., 0] -= mean[0]
    img_array[..., 1] -= mean[1]
    img_array[..., 2] -= mean[2]
    
    return np.expand_dims(img_array, axis=0)

# Run inference
image_path = "path/to/your/lesion_image.jpg"  # Change this
img_array = preprocess_image(image_path)
prediction = model.predict(img_array)[0][0]

# Display result
print(f"\n{'='*50}")
print(f"Prediction: {'MALIGNANT' if prediction >= 0.5 else 'BENIGN'}")
print(f"Confidence: {prediction:.2%}")
print(f"{'='*50}")

Model Architecture

ResNet50 Backbone:

Pre-trained on ImageNet
Input: 224×224×3 RGB images
Preprocessing: ImageNet normalization (Caffe-style)

Custom Classification Head:

GlobalAveragePooling2D
    ↓
BatchNormalization
    ↓
Dropout (0.5)
    ↓
Dense (512, ReLU)
    ↓
BatchNormalization
    ↓
Dropout (0.5)
    ↓
Dense (256, ReLU)
    ↓
Dropout (0.3)
    ↓
Dense (1, Sigmoid)

Training Configuration

Loss: Weighted Binary Cross-Entropy (pos_weight=4.12 for class imbalance)
Optimizer: AdamW (learning_rate=1e-4)
Batch Size: 16
Epochs: 25
Data Split: 70% train / 15% validation / 15% test (lesion-level stratified)
Callbacks: ModelCheckpoint, EarlyStopping, ReduceLROnPlateau

Educational use only — NOT for real-world or clinical applications.

Medical & Regulatory Disclaimer

This model is for educational and research purposes only. It must not be used in real-world/clinical contexts (including screening, triage, diagnosis, or treatment). It is not a medical device and has not been reviewed or approved by any regulatory authority (e.g., FDA/MHRA). Model outputs must not replace clinical judgment; any suspicious lesions should be evaluated by a qualified clinician.

Scope & Limitations (optional to keep)

Trained on HAM10000 dermoscopic images; may not generalize to other image types, devices, clinics, or skin tones.
Binary prediction (malignant vs. benign); chosen threshold strongly affects precision/recall.

Dataset

HAM10000: Human Against Machine with 10,000 training images

10,015 dermoscopic images
7 lesion types collapsed into binary labels:
- Benign: nv (melanocytic nevi), bkl (benign keratosis), df (dermatofibroma), vasc (vascular lesions)
- Malignant: mel (melanoma), bcc (basal cell carcinoma), akiec (actinic keratoses)
Source: HAM10000 on Kaggle

Citation

If you use this model, please cite the HAM10000 dataset:

@article{tschandl2018ham10000,
  title={The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions},
  author={Tschandl, Philipp and Rosendahl, Cliff and Kittler, Harald},
  journal={Scientific data},
  volume={5},
  number={1},
  pages={1--9},
  year={2018},
  publisher={Nature Publishing Group}
}

License

MIT License - See repository for details.

More Information

GitHub Repository: https://github.com/devatreya/Skin-lesion-classification-using-CNN
Full Documentation: See GitHub README for complete training details
Model File: resnet50_best.h5 (214 MB)

Downloads last month: -; Downloads are not tracked for this model. How to track