MPhil Bone Imaging AI Workflow Supervisor Package

Prepared on: 2026-06-16

This ZIP is a supervisor-ready blueprint for a realistic MPhil project using orthopedic imaging AI. It covers femur/leg fractures, vertebral/spine analysis, modalities, public datasets, Pakistan hospital data acquisition, MONAI/BoneJ roles, compute options, limitations, timeline, and CLI execution.

What this package is

A detailed research-engineering plan for a reproducible medical imaging workflow.

What this package is not

It is not a clinical diagnostic product, not a trained model, and it does not include patient data.

Current runnable status

The CLI pipeline is runnable as a scaffold:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
bash scripts/run_pipeline.sh

If no real features are present, the training script creates a small DEMO feature table so the code path can be tested. Demo metrics are not research findings. For thesis work, replace data/processed/features.csv with patient-level, de-identified, real extracted features.

Recommended reading order

1. MASTER_FULL_DETAILED_WORKFLOW.txt
2. docs/00_supervisor_summary.md
3. docs/01_research_goals.md
4. docs/02_workflow_detailed.md
5. docs/03_datasets_and_access.md
6. docs/04_pakistan_data_acquisition.md
7. docs/05_compute_budget_cloud_cli.md
8. configs/workflow.json
9. diagrams/workflow_mermaid.md
10. docs/09_google_colab_test_guide.md
11. templates/hospital_data_request_letter.txt

Main recommendation

Start with public data to prove the workflow, then add 100–200 local Pakistan DICOM cases only after ethics approval and de-identification.

Hard boundary

Never upload identifiable patient data to public cloud or GPU marketplace providers without explicit institutional approval and proper safeguards.

Key weaknesses to address before thesis claims

1. Real feature extraction is still a stub; DICOM manifesting exists, but morphometry/radiomics extraction must be implemented for the chosen modality.
2. Dataset choice must be narrowed to one primary path: simple X-ray classification, spine CT segmentation/morphometry, or local femur/hip workflow.
3. Labels and splits need a formal patient-level manifest to avoid leakage.
4. MONAI training is currently a placeholder, not a complete deep-learning experiment.
5. Colab is suitable for experiments, but not for storing identifiable data or running persistent jobs.

See REPRODUCIBILITY.md and GITHUB_UPLOAD_CHECKLIST.md before using Colab or GitHub.

Public-data Colab viability test

Use the UCI Vertebral Column public dataset to test the baseline without patient data:

python3 scripts/download_public_test_data.py --dataset vertebral_column --output data/processed/features.csv
bash scripts/run_pipeline.sh

For Colab-specific commands, see docs/09_google_colab_test_guide.md.

Public imaging datasets

For real imaging experiments, do not stop at the UCI smoke test. Use docs/10_public_dataset_strategy.md and:

bash scripts/download_public_imaging_data.sh rsna-cspine-kaggle data/raw/public_imaging/rsna_cspine

NIH ChestX-ray14 is useful for practicing large public X-ray workflows, but it is not an orthopedic fracture dataset. MURA, RSNA Cervical Spine, VerSe, CTSpine1K, and TCIA/NBIA are better aligned with bone/spine research depending on the final topic.