Breast Cancer Prediction ML Web App — Case Study

01 — Overview

What is MediPredict?

MediPredict is a breast cancer prediction web application that uses machine learning and deep learning to help detect the presence or absence of breast cancer at an early stage. The system accepts two types of input: clinical numerical data about the patient, or a histopathological microscopy image of breast tissue — giving users two independent prediction methods within a single application.

Breast cancer remains one of the leading causes of cancer death worldwide, with many cases going undetected until later stages due to diagnostic errors and limited access to specialist expertise. This project explored whether machine learning and deep learning could provide an accurate, accessible, and cost-effective alternative prediction tool — and deployed the best-performing algorithms into a usable web application anyone could interact with.

The project was completed as part of my BSc in Software Engineering at Cardiff Metropolitan University (via ICBT), achieving First Class Honours. The work covered the full pipeline: literature review, data acquisition, algorithm selection, training and evaluation, comparative analysis, system design using UML, web application development with Flask, form validation, and black-box testing.

Quick Facts

Context

BSc Software Engineering · Cardiff Met · 2022

Result

First Class Honours

Best ML Accuracy

97.14% — SVM on WBCD dataset

Best DL Accuracy

95.65% — CNN on BreakHis dataset

02 — The Problem

Why this matters.

⚠️

The Problem

Traditional breast cancer detection (mammography, ultrasound, clinical examination) depends heavily on specialist expertise and is prone to human error — with approximately 52% of diagnostic errors caused by misreading symptoms and 43% by overlooking them. Early-stage detection rates remain low in many countries, directly impacting survival rates.

✅

The Solution

A dual-method web application combining a trained SVM model (clinical data) and CNN model (tissue images) into a single accessible interface. Users input either their clinical measurements or upload a histopathological image — and the system returns a prediction instantly, with a clear message to consult a doctor before acting on results.

03 — Two Prediction Methods

Clinical data and image input — one app.

One of the distinctive features of this project is that it supports two completely independent prediction pathways in a single web application — each using a different dataset, a different class of algorithm, and a different input type.

Method I — Clinical Data

Machine Learning on WBCD Dataset

Users input 9 numerical clinical measurements (clump thickness, cell size uniformity, cell shape uniformity, etc.) from the Wisconsin Breast Cancer Diagnostic dataset. The trained SVM model predicts whether the tumour is malignant or benign. Dataset: 669 cases — 458 benign, 241 malignant. 70/30 train/test split with 10-fold cross validation.

97.14%

SVM accuracy
Best of 4 algorithms tested

Method II — Histopathological Image

Deep Learning on BreakHis Dataset

Users upload a microscopic biopsy image of breast tissue. The trained CNN model classifies the image as benign or malignant. Dataset: 9,109 images from 82 patients at 4 magnification levels (40×, 100×, 200×, 400×). Images resized to 224×224×3. 85% training / 5% testing / 10% validation split.

95.65%

CNN validation accuracy
Best of 3 algorithms tested

04 — Algorithm Comparison: Machine Learning

Testing four algorithms to find the best.

Four machine learning algorithms were trained and evaluated on the Wisconsin Breast Cancer Diagnostic dataset. All were measured on accuracy, precision, recall, and F1 score for both benign and malignant classes, as well as confusion matrix results.

Algorithm	Accuracy	Precision (B/M)	Recall (B/M)	Result
SVM Winner	97.14%	0.98 / 0.96	0.98 / 0.96	204/210 correct
KNN	95.71%	0.96 / 0.95	0.97 / 0.94	2nd place
Naïve Bayes	95.23%	—	—	3rd place
CART	90.47%	—	—	Lowest performer

SVM was selected for deployment. It correctly classified 204 out of 210 test cases — 130 benign correctly identified as benign, 74 malignant correctly identified as malignant, with only 6 misclassifications (3 false positives, 3 false negatives). This result is consistent with other published research showing SVM accuracy between 96.6% and 97.9% on the same dataset.

05 — Algorithm Comparison: Deep Learning

Testing three deep learning models on image data.

Three deep learning architectures were trained and evaluated on the BreakHis histopathological image dataset. Models were compared on validation accuracy and validation loss — the key metrics for assessing how well a model generalises to unseen data.

Model	Validation Accuracy	Validation Loss	Result
CNN Winner	95.65%	8.68%	Best on both metrics
DenseNet121	95.65%	16.78%	Equal accuracy, higher loss
ResNet50	94.78%	19.09%	Lowest performer

CNN was selected for deployment. While CNN and DenseNet121 achieved equal validation accuracy (95.65%), CNN's validation loss of 8.68% was significantly lower than DenseNet121's 16.78% — indicating CNN generalises better and produces more confident, reliable predictions on unseen images.

06 — Web Application

MediPredict — the deployed product.

The best-performing models (SVM + CNN) were integrated into a Flask web application called MediPredict. The interface was designed to be approachable and clear for non-technical users — including a disclaimer popup before predictions, validated input forms, and a results section that always reminds users to consult a doctor before acting on the output.

MediPredict homepage — "Let's Stand Together and Fight Breast Cancer"

Method I — clinical data input and prediction result

Method II — histopathological image upload and result

07 — Application Pages

A complete user-facing product.

Beyond the prediction tool, MediPredict includes an About Us page explaining the system and its purpose, a Preventions page giving users evidence-based guidance on reducing breast cancer risk, and thorough input validation on all forms.

About MediPredict

Preventions & awareness

08 — Tech Stack

Built with Python and Flask.

The entire stack is Python-based — scikit-learn and TensorFlow/Keras for model training, Flask to serve the models as a web application, and standard HTML/CSS/JS for the front end. The trained models were serialised and loaded at runtime, with Flask handling the routing, form submission, prediction calls, and response rendering.

🐍

Python

🧠

TensorFlow / Keras

📊

scikit-learn

🌐

Flask

🐼

Pandas / NumPy

📈

Matplotlib

09 — Outcomes

What was achieved.

97.14% SVM accuracy on clinical data 204/210 test cases correctly classified

95.65% CNN validation accuracy on images Lowest validation loss of the three models

1st First Class Honours BSc Software Engineering · Cardiff Metropolitan University

10 — Reflections

What I learned.

01

Algorithm selection is an empirical process. The literature suggested SVM and CNN would perform well — but implementing all the alternatives and measuring them directly was the only way to confirm this. The comparison process itself produced the most valuable learning, not just the final result.

02

Validation loss matters as much as validation accuracy. CNN and DenseNet121 achieved identical validation accuracy. It was the validation loss that differentiated them — CNN's significantly lower loss meant its predictions were more confident and its generalization more reliable. Accuracy alone is not enough to select a model.

03

ML in healthcare demands responsibility. Designing the disclaimer popup and the "consult your doctor" message wasn't just good UX — it was ethically necessary. A prediction tool in a medical context needs to be transparent about what it is and what it isn't, especially when its outputs might influence someone's decisions about their own health.

Breast Cancer —ML Prediction Web App