Prostate Radiomics ML

This repository contains the radiomics-only branch of the prostate cancer project. It keeps the feature extraction workflow for multiparametric prostate MRI, classical machine learning models, and deep tabular models trained only on extracted radiomics features.

It does not train neural networks on image slices, volumes, patches, CNN feature maps, image ViTs, or segmentation tensors.

Clinical Task

The main task is binary classification of clinically significant prostate cancer (csPCa), defined as ISUP >= 2, from radiomics features extracted from:

• T2W
• ADC
• DWI/HBV

The expected modeling label is label:

• 0: not clinically significant
• 1: clinically significant

Repository Layout

├── artifacts/
│   ├── data.csv                         # Cohort table with paths, metadata, and labels
│   └── radiomics/                       # Radiomics feature CSVs
├── data_analysis/                       # Cohort and image descriptive notebooks
├── data_structuring/                    # Initial cohort table assembly
├── results/
│   └── radiomics/                       # Radiomics model outputs
├── train/
│   ├── common/                          # Shared data and runtime utilities
│   └── radiomics/
│       ├── 1_extract_radiomics/         # PyRadiomics extraction code and parameters
│       ├── 2_modeling/                  # Classical ML, comparisons, and DL runners
│       └── deep_models/                 # Modular deep tabular model definitions
├── requirements.txt                     # Base extraction and classical ML pipeline
└── requirements-deep-radiomics.txt      # Optional TensorFlow deep tabular models

Installation

Base pipeline:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Deep tabular radiomics models:

pip install -r requirements-deep-radiomics.txt

Main Workflow

1. Extract Radiomics Features

python train/radiomics/1_extract_radiomics/extract_radiomics.py

This writes one CSV per modality and spatial region under artifacts/radiomics/.

2. Concatenate Modalities

python train/radiomics/2_modeling/0_build_concatenated_feature_table.py \
  --radiomics_root artifacts/radiomics \
  --mode gland \
  --keep_shape_from t2 \
  --output artifacts/radiomics/concatenated_data/features_all_gland.csv

3. Train Classical ML Models

python train/radiomics/2_modeling/1_train_and_evaluate.py \
  --csv features_all_gland.csv \
  --data_pre artifacts/radiomics \
  --results_base results/radiomics \
  --feature_strategy most_discriminant \
  --n_splits 5 \
  --n_repeats 10

4. Train Deep Tabular Models

The deep models receive only the radiomics feature matrix. Feature selection, imputation, and scaling are fitted only on the training split to avoid leakage.

Available architectures:

• transformer
• capsnet
• transformer_capsnet

python train/radiomics/2_modeling/4_train_tabular_transformer.py \
  --csv features_all_gland.csv \
  --data_pre artifacts/radiomics \
  --output_dir results/radiomics/deep_tabular_models \
  --run_name gland_transformer \
  --architecture transformer \
  --feature_selection most_discriminant \
  --epochs 300 \
  --batch_size 16 \
  --patience 50

Run the full deep tabular suite:

python train/radiomics/2_modeling/4_run_deep_tabular_suite.py \
  --csv features_all_gland.csv \
  --data_pre artifacts/radiomics \
  --output_dir results/radiomics/deep_tabular_models \
  --run_prefix final_5fold \
  --architectures transformer capsnet transformer_capsnet

Deep Model Implementation

Deep model code is modularized under train/radiomics/deep_models/:

• config.py: shared DeepTabularConfig
• layers.py: positional embedding, attention pooling, capsule routing, capsule length, Transformer blocks
• losses.py: binary focal loss and CapsNet margin loss
• architectures.py: model builders and architecture-specific target/prediction helpers

The transformer architecture extends the reference tabular Transformer with a residual dense encoder, LayerNormalization, semantic radiomics tokens when feature names are available, token dropout, and AdamW. By default it uses binary focal loss with gamma=2.0 and alpha=0.35; use --transformer_loss bce to run a binary cross-entropy ablation with balanced class weights.

The capsnet architecture matches the reference CapsNet notebook and uses CapsNet margin loss with one-hot targets, capsule lengths as outputs, routing_iterations=3, CategoricalAccuracy, AUC, and balanced class weights during training.

The transformer_capsnet architecture remains the project hybrid benchmark and uses a binary probability output with focal loss.

Methodological Safeguards

• Patient-level splitting by patient_id.
• Feature selection is performed inside each training split or fold.
• Imputation and scaling are fitted only on training data.
• Classical ML models share the same fold plan for fair comparison.
• Deep tabular models can reuse predefined folds and shared fold-wise feature plans.
• Deep learning is restricted to tabular radiomics variables.

Additional Documentation

The detailed radiomics workflow is documented in train/radiomics/README.md.