Skin Disease Detection using CNN

The Skin Disease Detection system is a healthcare-AI final-year project that identifies common skin conditions from ordinary smartphone photos. Dermatologists are scarce — India has roughly one dermatologist per 100,000 people — so an AI triage assistant that runs on a phone camera is both socially relevant and technically rich. This skin disease detection using CNN project trains a convolutional neural network on a HAM10000-style dermoscopic dataset covering seven lesion categories (melanoma, basal cell carcinoma, actinic keratosis, benign keratosis, dermatofibroma, vascular lesions, melanocytic nevi) and serves it through a mobile-friendly Flask web app called DermaAI.

Upload a photo and DermaAI returns the top-3 predicted conditions with confidence scores, plain-language condition information and general care guidance — the kind of end-to-end demo that holds an examiner's attention from the first upload. The report documents the full ML pipeline: dataset analysis, class-imbalance strategy, transfer learning, augmentation, confusion-matrix evaluation and a responsible-use disclaimer section that shows the maturity examiners reward.

This is a built-to-order software project: code, trained model weights and the complete viva kit (report, PPT, Q&A) are prepared fresh for you.

1. The user takes or uploads a close-up photo of the skin lesion in the DermaAI app.
2. The preprocessing pipeline crops, resizes and normalizes the image (224×224) for the CNN.
3. The fine-tuned MobileNetV2 model classifies the lesion into one of seven categories.
4. DermaAI shows the top-3 predictions with confidence bars, plus an info card for the top match: what the condition is, typical signs and general care guidance.
5. A Grad-CAM overlay highlights the lesion region the model focused on, and high-risk predictions (e.g. suspected melanoma) carry a clear prompt to consult a dermatologist.

• Multi-class classification of 7 common skin conditions from phone photos
• Top-3 predictions with confidence scores and plain-language condition info
• General care guidance and "see a dermatologist" urgency flags per condition
• Mobile-first DermaAI web app — works from any smartphone browser
• MobileNetV2 transfer learning, optimized for fast on-device-style inference
• Class-imbalance handling for rare but critical classes like melanoma
• Grad-CAM visualization showing which lesion area drove the prediction
• Condition info cards: description, typical symptoms, general care tips
• Training notebooks with accuracy/loss curves and per-class precision/recall
• Photo-capture guidance in-app (lighting, focus, distance) for better results

• Complete source code (training, evaluation, inference, Flask app)
• Trained model weights (.h5, plus TensorFlow Lite option)
• Dataset preparation and augmentation scripts
• Project report PDF (dermatology background, methodology, results, ethics/disclaimer)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, multi-class metrics, Grad-CAM, class imbalance)
• Setup guide (environment, dependencies, dataset download steps)

FAQs

1. Why is skin disease detection a strong final-year AI project? It combines a real access problem (dermatologist shortage) with a full multi-class deep-learning pipeline — transfer learning, class imbalance, augmentation, Grad-CAM and mobile deployment — giving examiners many angles to probe and you many angles to shine.
2. What dataset is the model trained on? A HAM10000-style dataset of 10,000+ dermoscopic images across seven diagnostic categories, released by the ISIC archive community. The report documents class distribution and the imbalance strategy used for critical classes like melanoma.
3. How does the app handle the melanoma risk responsibly? High-risk predictions display a prominent "consult a dermatologist promptly" banner, and every result screen carries the screening-aid disclaimer. The report's ethics section discusses false negatives explicitly — a discussion point that impresses examiners.
4. Will it work on my phone for the demo? Yes — DermaAI is a mobile-responsive web app: open the URL on your phone, snap or upload a lesion photo, and get top-3 predictions in under a second. No app-store install needed for the review.
5. Do I need a GPU? Only for training (scripts and cloud-GPU guidance included). The shipped trained weights infer in under a second on a regular laptop CPU.
6. How accurate is it on real phone photos? Design target: ~85–90% top-1 accuracy on held-out test images after training — the report documents the evaluation procedure to run during your build. Casual phone photos with good lighting should perform close to that, while poor lighting degrades results — the report's error analysis documents this honestly and the app gives capture tips.