Alzheimer's Detection using MRI

Alzheimer's disease begins damaging the brain years before memory symptoms appear, and structural MRI can reveal its fingerprints — hippocampal shrinkage, ventricular enlargement, cortical thinning — long before a clinical diagnosis. But measuring those changes by eye is subjective, and early stages like mild cognitive impairment (MCI) overlap heavily with normal aging, which is why automated staging is an active research area. The Alzheimer's Disease Neuroimaging Initiative (ADNI) has collected standardized longitudinal MRI for this exact purpose, making it the canonical academic dataset for the task. This project builds the standard student approach: a 3D CNN (with a slice-based 2D alternative) that stages scans into AD, MCI or CN, trained with strict subject-wise splitting — the methodological point examiners probe first, since random slice splits silently leak and inflate accuracy. It demonstrates neuroimaging ML end to end while being explicit that it is a research staging aid, never a diagnosis.

1. T1-weighted brain MRI scans are sourced from ADNI (via approved LONI IDA access) or from the openly downloadable OASIS 2D slice set for the quick-start path, with labels AD, MCI or CN.
2. Volumes are skull-stripped, registered to a common template, intensity-normalized and resampled; the 2D path extracts mid-temporal axial slices.
3. Subjects — never individual slices — are split into train/validation/test sets, so no brain appears in more than one split.
4. A 3D CNN (or 2D EfficientNet on slices with majority voting per subject) trains with augmentation — flips, rotations, elastic deformations — logging loss and per-class metrics per epoch.
5. The evaluation notebook computes final metrics on the held-out subjects: 3-class accuracy, per-class precision/recall/F1, 3x3 confusion matrix and AD-vs-CN ROC-AUC, plus Grad-CAM volumes.
6. The Flask demo stages an uploaded scan, displays per-class probabilities with the hippocampal ROI overlay and biomarker panel, and labels the output as a research staging aid.

Tech stack:

• Python 3.10, PyTorch (3D CNN and 2D EfficientNet training and inference)
• NiBabel, MONAI (NIfTI loading, registration, skull-stripping transforms)
• scikit-learn (metrics, confusion matrix, subject-wise splits)
• 3D Grad-CAM implementation (explainability volumes)
• Matplotlib, Seaborn, Nilearn-style plotting (curves, overlays, matrices)
• Flask demo web app (scan upload, staging panel, ROI viewer)
• Jupyter notebook (buyer-run training and evaluation)

• [3-class MRI staging classifier] 3D CNN on T1-weighted volumes (with a lighter 2D slice-based EfficientNet alternative) predicting AD, MCI or cognitively normal.
• [Subject-wise split enforcement] Splitting utilities that keep every scan of one subject in a single fold — the mandatory guard against slice-leakage inflated accuracy, stated and tested in the notebook.
• [Hippocampal ROI visualization] The demo overlays the medial-temporal region of interest on the scan, showing the atrophy pattern the model keys on across CN, MCI and AD cases.
• [Biomarker feature panel] Optional volumetric features (hippocampal volume, ventricular volume, cortical thickness estimates) shown alongside the CNN prediction for interpretability.
• [Grad-CAM brain overlays] 3D Grad-CAM visualizations highlighting the brain regions behind each staging decision, included in the report as explainability evidence.
• [MCI-focused error analysis] The notebook breaks down AD↔MCI and MCI↔CN confusion specifically — the genuinely hard boundaries — instead of hiding behind a single accuracy number.
• [ADNI + OASIS data paths] Full pipeline documented for ADNI (approved access via LONI IDA) with an OASIS-based quick-start path using openly downloadable 2D axial slices.
• [Staging demo UI] The Flask app accepts an uploaded scan, returns staged probabilities per class, the ROI overlay and a plain-language note that this is a research aid, not a diagnosis.

• Complete source code (preprocessing, 3D CNN, 2D slice pipeline, Grad-CAM, demo app, evaluation utilities)
• Jupyter training and evaluation notebook (buyer-run: preprocess, subject-wise splits, train, evaluate, error analysis)
• Flask demo application with scan upload and staging panel
• Project report PDF (background, ADNI/OASIS analysis, leakage-safe methodology, evaluation, MCI error analysis, limitations)
• PPT presentation for final review
• Viva Q&A preparation document (3D CNNs, subject-wise splitting, MCI difficulty, Grad-CAM, biomarkers, why not a diagnostic device)
• Setup guide (environment, ADNI access steps, OASIS quick-start, training on CPU/GPU, running the demo)