csPCA

A modular Python project for prostate cancer segmentation focused on identifying clinically significant prostate cancer (csPCA) lesions from MRI data.

This repository combines:

• preprocessing utilities for image/label mapping
• model definitions for segmentation pipelines
• training code with custom loss functions
• configuration for dataset paths and training hyperparameters
• a classical ML exploratory workflow under csPCA_classical_ml/

Overview

The project is designed around a segmentation workflow for MRI-based lesion detection. It uses a configuration-driven setup to:

• load MRI and annotation data
• align imaging volumes with label masks
• prepare 2D slices or processed patches for training
• train a segmentation model with class imbalance handling
• evaluate using thresholds and post-processing

The codebase is organized to support experimentation with different model and training strategies while keeping dataset and pipeline configuration centralized in config.py.

Repository structure

csPCA/
├── config.py                     # Global project configuration and training settings
├── Documentation                 # High-level project notes and dataset description
├── preprocessing/               # File loading and label mapping utilities
│   ├── loading_files.py
│   └── mapping_labels.py
├── models/                      # Deep learning segmentation models
│   ├── boundarylossmodel.py
│   ├── boundarylossmodel_2p5d.py
│   ├── boundarylossmodel_multimodal.py
│   └── __init__.py
├── training/                     # Training utilities and loss functions
│   ├── losses.py
│   └── __init__.py
├── csPCA_classical_ml/          # Classical ML / image processing experiments
│   ├── Imageprocessing.py
│   ├── imgpro.py
│   └── main.py
├── .gitignore
└── README.md

Dataset assumptions

The project expects MRI and label data structured around the paths declared in config.py:

project_root = Path(__file__).parent
checkpoint_dir = project_root / 'checkpoints'
labels_root = project_root / 'data' / 'picai_labels'
mri_root = project_root / 'data' / 'mri_images'

This means the repository is intended to be used with local data folders such as:

data/
├── mri_images/
│   └── ...
└── picai_labels/
    └── ...

The configuration also sets training split ratios and class imbalance handling for positive and negative cases, making this workflow suitable for lesion detection datasets with imbalanced labels.

Key features

• MRI-based segmentation pipeline for prostate cancer
• class-imbalance-aware training configuration
• custom combined loss setup with dice, focal, and boundary terms
• augmentation and weighted sampling support
• checkpoint and early-stopping-oriented training workflow
• modular preprocessing and model components

Setup

1. Clone the repository.
2. Create a Python environment and install the required dependencies (for example: PyTorch, NumPy, scikit-image, and any image-processing libraries used by the project).
3. Prepare your local dataset and update the paths in config.py to match your environment.
4. Run the relevant training or inference entrypoint for your workflow.

Configuration

The main settings are defined in config.py and include:

• dataset paths
• image target size
• batch size and workers
• training epochs and learning rate
• augmentation options
• loss weights and class imbalance parameters
• threshold optimization and post-processing settings

Before training, review the values in config.py and adjust them for your dataset size, hardware availability, and experiment requirements.

Usage notes

This repository is organized as a research/development codebase rather than a single packaged CLI. The workflow is intended to be adapted to the local dataset and model experiments you want to run.

Typical workflow:

1. verify dataset paths in config.py
2. run preprocessing to align images and labels
3. execute training code using the model and loss definitions
4. inspect checkpoints and evaluate segmentation performance