Custom EfficientNet BEAST v2 - Skin Disease Detector AI

This repository contains a custom, high-performance EfficientNet-style model for the classification of 8 common skin conditions. This model was built from scratch and trained on a massively augmented dataset of over 33,000 images using an NVIDIA A100 80GB GPU.

Model Description

The model is a deep convolutional neural network based on the principles of EfficientNet (MBConv blocks, Squeeze-and-Excitation, compound scaling) but with a custom, up-scaled architecture designed for this specific task. It has been scaled to ~150 million parameters to effectively learn from a large-scale, high-resolution medical imaging dataset.

Model Architecture

• Model: Custom EfficientNet BEAST v2
• Parameters: ~150 Million
• Input Size: 512x512x3

Classes

The model is trained to classify the following 8 conditions:

• Acitinic Keratosis
• Basal Cell Carcinoma
• Dermatofibroma
• Nevus
• Pigmented Benign Keratosis
• Seborrheic Keratosis
• Squamous Cell Carcinoma
• Vascular Lesion

(Note: The 'Melanoma' class from the original dataset was not present in the processed data and is therefore not included in this version.)

🏆 Performance

This model achieved the following state-of-the-art performance on the validation set:

• Validation Accuracy: 88.96%
• Top-3 Accuracy: 97.89%
• Validation Loss: [Add your final Validation Loss here]

How to Use

You can easily load and use this model for inference using TensorFlow/Keras and the huggingface_hub library.

First, install the necessary libraries:

pip install tensorflow huggingface_hub
import tensorflow as tf
from huggingface_hub import from_pretrained_keras
import numpy as np
from PIL import Image

# Load the model from your Hugging Face Hub repo
model = from_pretrained_keras("Arko007/skin-disease-detector-ai")

print("Model loaded successfully!")

# Define the class names in the correct order
CLASS_NAMES = [
    'Acitinic Keratosis', 'Basal Cell Carcinoma', 'Dermatofibroma', 'Nevus', 
    'Pigmented Benign Keratosis', 'Seborrheic Keratosis', 
    'Squamous Cell Carcinoma', 'Vascular Lesion'
]

def preprocess_image(image_path, img_size=512):
    """Loads and preprocesses an image for the model."""
    img = Image.open(image_path).convert('RGB')
    img = img.resize((img_size, img_size))
    img_array = np.array(img)
    img_array = np.expand_dims(img_array, axis=0)
    img_array = tf.cast(img_array, tf.float32) / 255.0
    return img_array

def predict(image_path):
    """Runs inference on a single image."""
    processed_image = preprocess_image(image_path)
    predictions = model.predict(processed_image)
    predicted_class_index = np.argmax(predictions, axis=1)[0]
    predicted_class_name = CLASS_NAMES[predicted_class_index]
    confidence = np.max(predictions, axis=1)[0]
    
    print(f"Predicted Class: {predicted_class_name}")
    print(f"Confidence: {confidence:.2%}")
    return predicted_class_name, confidence

# Example usage:
# predict("path/to/your/skin_image.jpg")

## Training Data
The model was trained on the "Multiple Skin Disease Detection and Classification" dataset. The original dataset of ~3,300 images was augmented 10x using the `imgaug` library to create a final training set of over 33,000 images, which helped in achieving high generalization performance.

## Author
This model was built by Arko.