About BPAD Interactive
BPAD — Biomedical Physics with Applications to Disease — is a textbook project that connects the physics of measurement to the statistics of inference in clinical research. Its data-modeling chapter introduces the machine-learning workflow through a single running example: kidney tumour CT, segmentation masks, and the outcomes that follow surgery.
The chapter's ideas were originally illustrated with R/Shiny applications. Shiny requires a running R server, which makes the demonstrations awkward to share, slow to load, and impossible to use offline. This site reimplements each of those applications as a browser-native module: the simulation, the model fitting and the plotting all run in your own tab, in TypeScript, with no backend of any kind. Nothing you enter is transmitted anywhere.
How the modules are organized
The catalog follows eight clusters that trace the arc of an imaging-AI study.
Foundations & Study Design
Sections 1-3Estimands, dependence structure, acquisition shift, leakage, and the design decisions that determine whether a model result means anything at all.
Imaging Features & Preprocessing
Sections 4-6From CT volumes and masks to quantitative features: physical units, texture discretization, augmentation, and the physics of image formation.
Supervised Learning
Section 7Classification and regression on the kidney cohort: thresholds, ROC geometry, neighborhood methods, model comparison, and decision cost.
Time-to-Event & Longitudinal
Section 8Censoring, competing risks, landmark analysis, tumor growth kinetics, and joint models for repeated biomarker measurements.
Unsupervised Learning & Segmentation
Section 9Phenotyping, cluster stability, intensity and spectral segmentation, and the metrics used to judge a delineation.
Deep Learning & Image-Native AI
Section 10Capacity and overfitting, convolution kernels as physical operators, receptive fields, parameter budgets, and multimodal fusion.
Evaluation, Calibration & Uncertainty
Section 11Discrimination versus calibration, decision curves, conformal prediction intervals, and reproducibility auditing.
Translation, Fairness & Deployment
Sections 12-13Subgroup fairness, counterfactual explanation, shortcut learning, differential privacy, human-AI teaming, and model documentation.
How the simulations work
- Reproducibility. Every module uses a seeded pseudo-random generator. The same seed and the same control settings always produce the same numbers, so a figure can be recreated exactly in a lecture, a lab report or an exam.
- Real algorithms. Logistic and linear models are fit by iteratively reweighted least squares and QR-free normal equations; Kaplan–Meier curves, log-rank tests, ROC and precision–recall curves, k-means, hierarchical clustering, PCA, GLCM texture features and Otsu thresholding are all implemented directly rather than approximated.
- Synthetic but faithful data. The cohorts are simulated with generative mechanisms that mirror the published KiTS-style kidney tumour data: skewed volumes, correlated radiomic features, informative censoring and protocol-dependent measurement error. No patient data is distributed with this site.
- Accessibility. Figures are SVG with text labels, controls are keyboard operable, and the colour palette is checked for contrast in both light and dark themes.
Credits and reuse
Built with the Statistics Online Computational Resource (SOCR) at the University of Michigan. The material is intended for teaching and self-study; you are welcome to use the modules in courses with attribution to the BPAD project and SOCR.