Medical Image Data Preprocessing for Anomaly Detection (Synthrad 2023 Pelvis)

This repository contains the modularized data preprocessing pipeline for the Synthrad 2023 Pelvis dataset. The pipeline transforms 3D CT/MR volumes into processed 2D slices, performing intensity normalization, automated artifact segmentation, and standardized geometric transformations.

---

Modular Component Overview

The logic is split between Driver Scripts (high-level orchestration) and Utility Modules (the processing engine).

1. Main Dataset Processing Scripts (Drivers)

These scripts handle the entire pipeline from loading to final export. Each exists in a sc_ (Single-Center) and mc_ (Multi-Center) version.

Script Name	Output Format	Input Slice Channels	Output Size	Description
{sc/mc}_dataset_processing_png.py	PNG (.png)	1 Slice (RGB)	(224, 224)	Extracts the current 2D slice, maps it to 'bone' colormap, and saves as a 3-channel PNG.
{sc/mc}_dataset_processing_nifti_rep.py	NIfTI (.nii)	1 Slice (Replicated)	(224, 224)	Extracts the current slice and replicates it across 3 channels into a NIfTI file.
{sc/mc}_dataset_processing_nifti_con.py	NIfTI (.nii)	3 Consecutive Slices	(224, 224)	Extracts (prev, current, next) slices to form a 3-channel NIfTI, capturing spatial context.

2. Utility Modules (utils/)

Module	Core Responsibility
artifact_detector.py	Houses the MetalArtifactDetector class for HU-based thresholding and MR refinement.
processing_utils.py	Manages the geometry pipeline: padding, resizing, cropping, and min-max normalization.
io_utils.py	Standardized saving/loading for NIfTI and PNG to ensure consistent data types.
path_utils.py	Handles the automated creation of the img, label, and bodymask subfolder structures.

---

The Preprocessing Pipeline

All driver scripts follow a standardized 5-step pipeline:

1. Loading: Pairs MR, CT, and Mask volumes for a specific Patient ID.
2. Normalization: Applies a body mask to MR data and performs min-max normalization ($0 \rightarrow 255$) for consistent intensity ranges.
3. Anomaly Detection: For "Ungood" (abnormal) scans, the system calculates a dynamic threshold ($\tau$) based on CT Hounsfield Units to isolate implants, which is then refined using the MR signal.
4. Geometric Standardization:
   • Square Padding: Slices are center-padded to a 1:1 aspect ratio.
   • Resizing: Slices are interpolated to a base size of 240x240.
   • Center Crop: A final crop to 224x224 is performed to remove edge artifacts.
5. Structured Export: Slices are sorted into train, valid, or test directories. A specialized "Whole Patient" export is performed for test cases to allow for full volumetric evaluation.

---

Usage

Run the driver script corresponding to your required center type and output format.

Single-Center PNG Example:

python sc_dataset_processing_png.py \
  --dir_pelvis /path/to/raw/data/Task1/pelvis \
  --dir_output /path/to/save/png_dataset