Skin Cancer Detection

About the skin cancer detection M.Tech project

Melanoma and other skin cancers are far easier to treat when a suspicious mole is checked early, yet most villages have no dermatologist. This skin cancer detection M.Tech project puts the first check inside an Android phone. The user photographs a skin spot, and an int8 model running on the phone sorts it into one of the seven HAM10000 lesion types, shows the top three results with a heatmap and gives a simple risk level.

The research question: can a model small enough for a low-cost phone keep its accuracy after integer quantization and still catch rare, dangerous classes such as melanoma?

Research behind it

Base paper. The study starts from A Novel End-to-End Deep Convolutional Neural Network based Skin Lesion Classification Framework, published in Expert Systems with Applications (Elsevier, 2024). It adds segmentation and feature extraction steps, classifies seven HAM10000 lesion types with S-MobileNet, and shows that a compressed S-MobileNet can even beat the uncompressed one.

Research gap. Compression in a paper is not the same as an int8 TFLite file on a real handset, where size, latency and memory decide whether the app is usable. HAM10000 is dominated by benign moles, so overall accuracy can hide poor melanoma recall. Quantization-aware training and post-training quantization are not compared for this task, and screening needs a path from the phone to a doctor.

Proposed system and novelty. MobileNetV3-Small is fine-tuned with focal loss and class-balanced sampling, so rare classes count more during training. It is then trained again with quantization-aware training, which lets the network adapt to int8 before conversion. A lesion-wise split keeps photos of the same lesion out of both train and test sets. Latency, memory and size are measured on a real Android phone.

Models compared. MobileNetV2, EfficientNet-Lite0, a full-precision MobileNetV3-Small and the base paper S-MobileNet re-implemented, against the proposed QAT model, plus float16 and int8 post-training versions. The ablation swaps focal loss for cross-entropy, QAT for post-training quantization and removes balanced sampling.

The same small-model idea is used in the lightweight plant disease detection project, which is a good comparison if your guide prefers agriculture.

How it works

• The user photographs the spot with an on-screen guide to keep it centred.
• The app checks blur and lighting and asks for a retake if the photo is poor.
• The int8 TFLite model predicts the top three classes on the phone.
• A light heatmap shows the region the model used.
• The app maps the class to a risk level and shows plain advice.
• When online, the scan syncs to the Flask server for doctor review.

Users and modules

There are three roles. The user scans spots, keeps a history with photos, sets reminders and downloads a PDF summary. The doctor works through a review queue of shared scans, adds remarks and marks referrals. The admin manages users, checks the dashboard, views the model comparison page and uploads a new model file. The Android app is written in Java, and the server side is Flask with MySQL.

What you get

• Phase 1 Research: Base paper study: S-MobileNet end-to-end framework (Expert Systems with Applications, 2024); Literature survey of 15-20 recent skin lesion papers; Research gap, novelty and objectives; HAM10000 EDA notebook with class imbalance study; Synopsis and Phase-1 report.
• Phase 2 Implementation: Android app (Java) with offline TFLite prediction; Flask API, doctor and admin panels; Baselines and S-MobileNet re-implementation; MobileNetV3-Small with focal loss; PTQ and QAT int8 TFLite models; MySQL database and trained models.
• Results: Accuracy, macro F1 and melanoma recall; Size, latency and memory on a real phone; Ablation of loss and quantization method; results.xlsx and all graphs in high resolution.
• Thesis and paper: Thesis written for you (about 90 pages); IEEE conference paper draft; Journal paper draft; Publication guide.
• Reviews, viva and support: PPTs for Review 1, Review 2 and final viva; Viva book with 50+ questions; A1 poster and project diary; Demo and code walkthrough videos; 1-on-1 setup, 2 revisions and support till viva.

Who can use this project

M.Tech and M.E. students of CSE and IT who want a healthcare dissertation with a mobile part will find this a good fit. It also suits anyone looking for a HAM10000 skin lesion classification project, a quantization-aware training TFLite Android project or a focal loss imbalanced medical imaging project. Every file has short comments so you can explain it in the viva.