BME5150 Chap.1 Medical AI Data and Radiomics
Medical AI Data and Radiomics
Start with the data-generating process
Medical artificial intelligence is built from measurements created by clinical workflows. Images depend on acquisition hardware, protocol and reconstruction. Labels depend on readers, tests or later outcomes. Patient variables arrive at different times and for different reasons.
A model can learn any regularity in that chain, including a site artefact that has nothing to do with disease. The chapter therefore begins before algorithm choice.
Radiomics exposes every stage
A radiomics workflow moves from image acquisition to segmentation, feature extraction, model fitting and evaluation.
Shape, intensity and texture descriptors make an image available to conventional machine-learning methods, but each transition adds assumptions. A segmentation boundary controls which voxels enter the calculation. Different filters and reconstruction settings change feature values. Redundant features can make selection unstable.
A valid performance claim belongs to the complete pipeline.
Heterogeneity needs a location
The course materials emphasise data heterogeneity in acquisition, segmentation and processing. Those sources require different responses. A scanner shift can be tested through feature stability and external evaluation. Annotation disagreement can be measured by comparing readers and reference procedures.
A change in prevalence alters precision even when sensitivity and specificity remain similar. Calling all of them domain shift loses the mechanism needed for a repair.
Privacy and transport are separate questions
Federated learning lets sites contribute to training without pooling raw patient records. It changes where information travels and may support collaboration across institutions.
It does not remove site differences or create independent validation. A federated model can still learn local artefacts and still needs testing on cases excluded from every training update.
End with an auditable claim
State the target population, available inputs, outcome and intended decision before fitting. Keep evaluation cases outside transformations and tuning.
When reporting performance, name the data boundary and the most likely reason the pipeline could fail elsewhere. That sequence turns a technical result into a clinical evidence statement.
What this chapter covers
- 01
Medical data types and intended use
- 02
Radiomics from images to quantitative features
- 03
Acquisition and segmentation heterogeneity
- 04
Data augmentation and simulation
- 05
Federated learning and privacy boundaries
- 06
Internal and external validation
Worked example · free
Locate the first invalid pipeline decision
- 3Identify information that crossed the evaluation boundary.
- 2Explain why a random two-hospital split is not necessarily external.
- 3State a valid development and evaluation design.
Key terms
- Radiomics
- Radiomics is the extraction of quantitative descriptors from medical images for modelling and analysis.
- Segmentation
- Segmentation is the delineation of an anatomical or pathological region used for measurement or prediction.
- Data Heterogeneity
- Data heterogeneity is variation in acquisition, patients, annotation or processing across records or sites.
- Federated Learning
- Federated learning trains across sites by exchanging model updates rather than pooling raw patient data.
- External Validation
- External validation applies a locked pipeline to a meaningfully separate population, site or time period.
Medical AI Data and Radiomics FAQ
Why can a model learn the hospital instead of the disease?
Hospital-specific scanners, protocols, reconstruction and patient pathways can correlate with the outcome. A flexible model may use those easier signals. Hold out sites, inspect feature distributions and report site-wise errors to test whether apparent disease prediction depends on institutional artefacts.
Does federated learning guarantee privacy and generalisability?
No. It reduces the need to pool raw patient records, but model updates can still require privacy safeguards, and site heterogeneity remains. Generalisability needs an independent evaluation boundary, even when development was distributed across several institutions.
What makes a validation set truly independent?
Its cases must not influence scaling, imputation, feature selection, tuning, threshold choice or model comparison. A stronger external design also separates the data-generating setting by site, time or population so that transport rather than row novelty is tested.
Assessment move
Draw the five-stage radiomics pipeline from memory and place one failure mechanism beneath each box. For acquisition, use protocol or scanner variation. For segmentation, use reader disagreement. For features, use instability or redundancy. For modelling, use overfitting. For evaluation, use leakage or weak transport.
Then take one published performance claim from your notes and rewrite it with a population, split boundary, metric and external limitation. When revising federated learning, keep two questions separate: where data travel and whether the trained pipeline generalises. Finish by explaining why augmentation creates more training instances but not more independent patients.
Add a counterfactual for each stage: state what would change in the result if only that stage changed. This forces the limitation to identify a mechanism instead of repeating a generic warning about medical data.
Working through Medical AI Data and Radiomics in BME5150? Sia is AskSia’s AI Computer Science tutor — ask any BME5150 Medical AI Data and Radiomics question and get a clear, step-by-step explanation grounded in how BME5150 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
