Breast Cancer Detection using CNN

Breast cancer diagnosis rests on the pathologist's microscope: H&E-stained tissue slides are examined at multiple magnifications for the architectural and nuclear changes that separate benign from malignant tissue. It is skilled, time-consuming work, and pathologist workloads keep growing — which is why automated histopathology analysis is one of the most published areas in medical computer vision. The BreaKHis dataset (Spanhol et al., 2016) — 7,909 microscopy images from 82 patients, labeled benign or malignant across 8 subtypes and 4 magnifications — is the standard academic benchmark for the task. This project builds the canonical student solution: a transfer-learned patch classifier with stain normalization, strict patient-wise splitting (the leakage trap examiners ask about first), per-magnification evaluation and a demo that visualizes what benign vs malignant tissue looks like. It is designed as a rigorous classification build while being explicit that it is a research prototype — histopathology diagnosis stays with the pathologist.

1. BreaKHis images are loaded (7,909 H&E patches, 700x460 px, benign vs malignant labels with 8 subtype labels and 4 magnification levels) and grouped by patient ID.
2. Patients are split into train/validation/test sets; patches inherit their patient's split, guaranteeing no leakage between sets.
3. H&E stain normalization is applied, patches are resized to 224x224 and augmented with rotations, flips and color jitter appropriate for microscopy.
4. An ImageNet-pretrained ResNet50 (or EfficientNet-B0) is fine-tuned with a binary head, logging loss, accuracy and AUC per epoch on the validation patients.
5. The evaluation notebook computes final metrics on held-out patients — accuracy, AUC, sensitivity, specificity, confusion matrix — broken down per magnification level and per subtype.
6. The Flask demo classifies an uploaded patch, shows the benign/malignant probabilities with the nuclei-highlight overlay, and labels the output as a research aid.

Tech stack:

• Python 3.10, PyTorch with timm (ResNet50 / EfficientNet-B0 transfer learning)
• OpenCV, scikit-image (patch extraction, stain normalization)
• scikit-learn (metrics, ROC-AUC, patient-wise splits)
• Matplotlib, Seaborn (curves, ROC plots, confusion matrix)
• Flask demo web app (patch upload, comparison viewer, nuclei overlay)
• Jupyter notebook (buyer-run training and evaluation)

• [Binary patch classifier] ResNet50 / EfficientNet-B0 fine-tuned on BreaKHis patches for benign vs malignant classification, with the 8-subtype experiment documented as an extension.
• [Patient-wise split enforcement] Splitting by patient ID (82 patients), never by image — patches from one patient are near-duplicates, and the notebook verifies the split is leak-free.
• [Stain normalization] Macenko-style H&E stain normalization in preprocessing, with a documented comparison of with/without normalization since stain intensity varies between labs.
• [Multi-magnification training] The model trains on the mixed 40x/100x/200x/400x set and the notebook reports per-magnification accuracy, showing which zoom levels are hardest.
• [Nuclei-highlight viewer] The demo can overlay detected nuclei contours on the patch, visualizing the nuclear crowding and irregularity the network keys on.
• [Subtype error analysis] Confusion analysis across the 8 BreaKHis subtypes (adenosis, fibroadenoma, phyllodes, tubular adenoma; ductal, lobular, mucinous, papillary carcinoma) for honest viva material.
• [Patch extraction pipeline] Whole-image to patch utilities with configurable patch size and stride, plus background filtering so empty slide regions don't pollute training.
• [Comparison demo mode] Side-by-side benign vs malignant patch viewer with magnification switching, built for classroom and viva demonstration.

• Complete source code (patch pipeline, stain normalization, training, demo app, evaluation utilities)
• Jupyter training and evaluation notebook (buyer-run: patient-wise splits, train, per-magnification evaluation, subtype analysis)
• Flask demo application with patch upload and comparison viewer
• Project report PDF (background, BreaKHis analysis, leakage-safe methodology, evaluation, subtype error analysis, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, patient-wise splitting, stain normalization, AUC, H&E imaging)
• Setup guide (environment, dataset download, training on CPU/GPU, running the demo)

• This is an academic prototype, not a medical device: not clinically validated, not certified, and must never guide diagnosis or treatment — histopathology diagnosis stays with the pathologist.
• BreaKHis comes from a single lab's 82 patients, so stain characteristics, scanner and population are narrow; performance on other labs' slides is not promised and stain normalization only partly bridges the gap.
• The base build classifies benign vs malignant only; the finer 8-subtype task is harder and is documented as an extension, not a solved problem.
• Patch-level predictions are not whole-slide diagnosis — real pathology reads the entire slide in context, which this project does not attempt.
• Class imbalance (5,429 malignant vs 2,480 benign) and the small patient count keep minority-subtype metrics noisy; the report states this openly.
• Reported metrics come from the buyer's own training run and vary with patient split, stain normalization and hyperparameters; design targets are goals, not guarantees.