BreastGuard AI Breast Ultrasound Image Classification with OpenCV & SVM (Benign / Malignant / Normal)

BreastGuard AI is a computer vision project that classifies breast ultrasound scan images into three categories — Benign, Malignant, or Normal. You upload an ultrasound image, OpenCV preprocesses it, and a tuned Support Vector Machine returns the class with a confidence score.

Why the Third Class Matters

Most breast cancer academic projects are binary — tumour present or absent. BreastGuard AI adds a Normal class, meaning the model must also recognise healthy tissue with no lesion at all. This three-way separation is significantly harder to train and is a strong talking point during viva, because it mirrors how a real screening workflow actually operates.

If your project requirement is classical ML on numeric lab values rather than image processing, MediPredict AI is the tabular alternative — it takes 30 cell-nucleus measurements as form input instead of a scan.

The Image Processing Pipeline

1. Grayscale conversion — ultrasound scans carry no useful colour information, so channels are collapsed to one
2. Resize to 128 × 128 — every input is normalised to a fixed dimension regardless of source resolution
3. Flattening — the 2D pixel matrix becomes a single vector of 16,384 features
4. StandardScaler normalisation — pixel intensities are centred and scaled before training
5. SVM classification — the tuned classifier predicts the class and outputs probability scores

Model Selection and Tuning

Three algorithms were trained and benchmarked on the same preprocessed data — Logistic Regression, K-Nearest Neighbors, and Support Vector Machine. SVM outperformed the others, and its kernel and regularisation parameters were then optimised using GridSearchCV, reaching 97% accuracy on the test split. The comparison results are included so you can present the full model-selection reasoning, not just the final number.

Key Features

• Three-class output — Benign, Malignant, and Normal, with per-class probability
• Direct image upload — drop in any ultrasound scan through the Flask interface
• Documented OpenCV pipeline — every preprocessing step is separable and inspectable
• Model comparison included — LR vs KNN vs SVM benchmark results ready to present
• Visual analysis — Matplotlib and Seaborn plots for confusion matrix and accuracy comparison
• Trained .pkl models included — run predictions immediately without retraining

Technology Stack

• Language: Python 3.10.11
• Backend: Flask
• Computer Vision: OpenCV
• Machine Learning: Scikit-learn (SVM, GridSearchCV, StandardScaler)
• Analysis & Plots: NumPy, Matplotlib, Seaborn
• Frontend: HTML5, CSS3, JavaScript

Complete Submission Package

• Full source code with the training notebook
• Trained model files (.pkl)
• Dataset structure and folder organisation guide
• Detailed project report in Word format
• Research paper draft, formatted for submission
• PowerPoint presentation
• Installation and setup support

Best Suited For

• Computer Vision and Medical Imaging final year projects
• Biomedical Engineering students working on diagnostic support systems
• Anyone who needs to demonstrate an end-to-end image classification pipeline — preprocessing through web deployment
• Students whose department expects a full documentation set alongside working code