Skin Lesion Classifier: A Deep Learning-Based Automatic Skin Lesion Recognition System

Skin Lesion Classifier is a deep learning web application based on Flask and TensorFlow/Keras, designed to provide an intelligent auxiliary tool for skin lesion screening. The system can automatically classify skin lesion images into 7 types (including malignant lesions like melanoma), with core features such as user authentication, image upload and analysis, AI classification prediction, result dashboard, and PDF report generation. This is an open-source project maintained by alokrk04, and the source code is available on GitHub (link: https://github.com/alokrk04/skin).

Skin cancer is one of the most common types of cancer globally, and early detection and diagnosis are crucial for improving the cure rate. With the development of deep learning technology, computer-aided diagnosis systems have shown great potential in the field of dermatology. This project aims to leverage these technologies to provide an efficient and accurate automatic skin lesion classification tool to assist medical professionals and researchers in screening work.

Skin Lesion Classifier is a complete web application that supports classifying skin lesions into 7 categories:

1. Actinic Keratosis and Intraepithelial Carcinoma (AKIEC): Precancerous lesions that require early identification to prevent malignant transformation;
2. Basal Cell Carcinoma (BCC): The most common type of skin cancer, with low malignancy but requiring timely treatment;
3. Benign Keratosis-Like Lesions (BKL): Including benign lesions such as seborrheic keratosis, which need to be distinguished from malignant ones;
4. Dermatofibroma (DF): A common benign tumor;
5. Melanocytic Nevus (NV): Benign hyperplasia, with some changes requiring vigilance for malignant transformation;
6. Pyogenic Granuloma and Hemorrhage (VASC): Vascular lesions;
7. Melanoma (MEL): The most dangerous type of skin cancer, where early detection is crucial.

Technical Stack:

• Backend: Flask 3.1.0 (lightweight Python web framework);
• Deep Learning: TensorFlow 2.20.0 + Keras 3.12.0 (using HDF5 format to store pre-trained models);
• Database: SQLAlchemy ORM (with two core models: User and Test);
• Frontend: Bootstrap + Jinja2 template engine.

Model Details:

• Input Preprocessing: Only supports PNG/JPG/JPEG formats, maximum single file size is 16MB, uniformly resized to 224x224 RGB images and normalized;
• Model Architecture: Convolutional Neural Network (CNN), including input layer, convolutional layers, pooling layers, fully connected layers, and output layer (7-class classification);
• Evaluation Metric: AUC-ROC (standard metric for medical image classification);
• Datasets: HAM10000 (10k+ expert-annotated images) and HMNIST (Harvard Medical School dataset).

Deployment Steps:

1. Clone the repository and enter the directory;
2. Create and activate a virtual environment;
3. Install dependencies;
4. Configure security keys and API keys;
5. Run the application.

Main Routes: Including login/registration (/login, /register), prediction page (/prediction), result viewing (/results), history records (/previous_tests), etc.

Security Considerations:

• Production Environment: Use a strong random SECRET_KEY (stored in environment variables), replace SQLite with PostgreSQL, enable HTTPS, rate limiting, CSRF protection, input validation, WSGI servers like Gunicorn;
• Data Privacy: User authentication and authorization, encrypted image storage, access log auditing, compliance with medical data regulations such as HIPAA/GDPR.

Limitations:

• For educational and research use only, cannot replace professional medical diagnosis;
• Prediction accuracy depends on image quality and lighting conditions;
• There is a model loading delay for the first prediction;
• Currently only supports single image processing;
• The model's generalization ability is affected by population, equipment, and shooting conditions.

Future Directions:

• Short-term: Batch image processing, WebSocket real-time prediction, advanced analysis dashboard;
• Mid-to-long term: Model interpretability (Grad-CAM), mobile application, DICOM integration, Docker deployment, CI/CD pipeline.

Medical Value:

• Improve screening efficiency and speed up skin lesion classification;
• Serve as a second opinion system to reduce missed diagnoses and misdiagnoses;
• Help remote areas access high-quality screening services;
• Increase the early detection rate of malignant lesions like melanoma.

Technical Value:

• Provide an end-to-end medical AI application example;
• Demonstrate best practices for production-ready machine learning applications;
• Modular architecture for easy expansion and customization.

Summary: Skin Lesion Classifier is an excellent application case of deep learning in the medical field, but AI systems should serve as auxiliary tools for doctors rather than replacements. Human-machine collaboration is the best path for medical diagnosis.