ML & Modeling

Step 6 – ML & Modeling: training models on those features

Once you have a clean table of features and labels, you can finally train models. In this step you apply machine-learning or deep-learning algorithms (for example logistic regression, random forests, CNNs, or MIL models) to answer a specific question: classification, prediction, grading, risk stratification, or similar tasks.

Technical name: ML & Modeling

What this is

Train models using image patches, features, and labels to:

• Classify regions or slides.
• Predict outcomes.
• Support triage or risk stratification.

Typical questions

• “Can a model flag suspicious areas on a slide?”
• “Can we predict recurrence risk from histology?”
• “Do image features improve beyond standard clinical variables?”

Common tasks

• Patch‑level classification (tumor vs normal, pattern, grade).
• Slide‑level predictions (risk score, biomarker status, response likelihood).
• Combine image features with clinical/genomic data.
• Train, validate, and tune models.

Core tools (examples)

• PyTorch / TensorFlow / Keras — deep learning frameworks.
• scikit‑learn — logistic regression, random forest, SVM.
• MONAI — medical imaging ML framework.
• TIAToolbox — WSI components (tiling, aggregating tile predictions).

Clinician mental model

Move from measurements and labeled examples to algorithms that generalize and make predictions on new slides.

Ready-to-use code

• ML-01: Classical + deep starts — Random Forest on tabular morphometrics (Path A) plus PyTorch ResNet transfer learning and inference (Path B) to cover both feature-based and end-to-end image modeling.