Introduction
Computed Tomography (CT) is a cornerstone of modern diagnostic medicine, yet the automated analysis of 3D CT data remains a significant technical challenge. Unlike MRI, which provides high soft-tissue contrast, CT images often suffer from lower spatial resolution and systematic noise. Traditionally, interpreting these scans requires substantial manual effort from expert radiologists for tasks such as registration (aligning scans to a common coordinate system) and segmentation (delineating specific organs).
AutoCT is introduced as an open-source, Docker-based pipeline designed to automate the end-to-end processing of 3D CT scans. By integrating diffeomorphic registration with atlas-based segmentation, it provides a reproducible framework for extracting localized features, significantly reducing manual bottlenecks in large-scale medical imaging research.
The AutoCT Pipeline
The framework is a modular Python-based system packaged in a Docker container to ensure maximum portability and reproducibility. The pipeline consists of seven core modules that transform raw data into actionable clinical metrics.
- Conversion: Transforms raw DICOM images into the standardized NIfTI format.
- Preprocessing: Standardizes image orientation, resamples voxels to a uniform size, and performs bias correction.
- Bone Stripping: A critical step for CT analysis; it uses intensity thresholding and Gaussian smoothing to remove dense bone structures, allowing subsequent algorithms to focus on soft tissue.
- Registration: Utilizes Advanced Normalization Tools (ANTs) to perform diffeomorphic mapping, ensuring transformations are smooth and invertible.
- Segmentation: Applies the inverse of the registration mapping to a standard anatomical atlas to automatically label regions of interest in the patient's original physical space.
- GeoMeasures: Calculates geometric properties such as volume and surface area for each segmented region.
- WarpStats: Analyzes the deformation field using the Jacobian of the mapping to provide statistical insights into anatomical deviations.
Spatial Normalization and Diffeomorphic Mapping
A key innovation of AutoCT is its management of the transition between "Physical Space" (the patient's actual dimensions) and "Normalized Space" (a standard coordinate system like the MNI template). This bridge allows researchers to perform population-level statistics while maintaining individual-level clinical relevance.
By prioritizing diffeomorphic transformations, AutoCT ensures that the "warping" process preserves the underlying anatomical topology. This mathematical rigor prevents digital "tearing" or "folding," maintaining biological plausibility when mapping a standard atlas onto a unique patient scan.
Results and Quantification
The pipeline was validated using CT scans from the RSNA Intracranial Hemorrhage Detection challenge. Using the Harvard-Oxford cortical and subcortical structural atlases, the system successfully parcellated brain CTs into 115 distinct regions.
Beyond simple volume extraction, the pipeline exports a rich feature set for downstream machine learning, including mean, standard deviation, and entropy of the Jacobian from the deformation field. The inclusion of Jupyter-based GUIs further makes these complex tools accessible to clinicians who may not be familiar with command-line interfaces.
Limitations and Future Work
While the current iteration focuses on brain CTs, the authors acknowledge that future versions will need validation for other complex anatomical areas, such as the lungs or abdomen. Future research will also focus on performing sensitivity analyses on the extracted features to build predictive models for acute diagnosis and long-term patient prognosis.
Summary updated