[Paper Review] Learning Based Segmentation of CT Brain Images: Application to Post-Operative Hydrocephalic Scans

Venkateswararao Cherukuri, Peter Ssenyonga+4 more|PubMed|Dec 11, 2017

Fetal and Pediatric Neurological Disorders35 references17 citations

TL;DR

This paper proposes a learning-based segmentation method for post-operative CT brain scans in hydrocephalus patients, using a supervised pixel-wise classification framework with class-specific dictionary learning to accurately distinguish brain tissue, cerebrospinal fluid (CSF), and subdural hematomas. The approach achieves state-of-the-art performance with lower computational cost and robustness to limited training data, enabling reliable volume estimation for clinical evaluation.

ABSTRACT

Experiments performed on infant CT brain images acquired from the CURE Children's Hospital of Uganda reveal the success of our method against the state-of-the-art alternatives. We also demonstrate that the proposed algorithm is computationally less burdensome and exhibits a graceful degradation against a number of training samples, enhancing its deployment potential.

Motivation & Objective

• To address the challenge of segmenting post-operative CT scans in hydrocephalus patients, where distorted anatomy and variable-intensity subdural hematomas confound traditional intensity-based segmentation methods.
• To develop a supervised pixel-level classification framework that learns class-specific dictionaries to effectively represent brain tissue, CSF, and subdural collections.
• To reduce computational and memory overhead compared to existing methods, enhancing deployability in low-resource clinical settings.
• To ensure robustness to variations in training data size, enabling practical use with limited annotated scans.
• To enable accurate volume estimation of brain and CSF for evaluating surgical outcomes in pediatric hydrocephalus.

Proposed method

• The method employs a dictionary learning framework that learns three class-specific dictionaries—one for brain tissue, one for CSF, and one for subdural hematomas—using labeled training data.
• It formulates segmentation as a sparse coding problem, solving for the sparsest representation of each image patch using the learned dictionaries to assign pixel labels.
• The framework incorporates a discriminative term that enhances separation between classes by emphasizing inter-class differences in the dictionary learning process.
• A customized training and testing procedure is designed to handle the unique challenges of post-op scans, including geometric distortion and intensity variability in subdurals.
• The method computes computational complexity and memory footprint analytically, showing significant reduction compared to baseline methods like DDLS and patch-based SRC.
• The final classification uses a linear classifier trained on sparse codes, enabling fast inference with minimal overhead.

Experimental results

Research questions

• RQ1 Can a supervised dictionary learning approach outperform traditional intensity-based segmentation in post-operative hydrocephalic CT scans with complex subdural pathology?
• RQ2 How does the proposed method handle the variability in subdural intensity and geometry across different patients?
• RQ3 What is the computational and memory efficiency of the proposed method compared to existing state-of-the-art techniques?
• RQ4 How does the method perform when trained on limited amounts of data, and does it degrade gracefully with fewer samples?
• RQ5 Can the method reliably separate subdural hematomas from brain tissue and CSF to enable accurate volume estimation for clinical evaluation?

Key findings

• The proposed method achieves state-of-the-art segmentation accuracy on post-operative CT scans from CURE Children's Hospital of Uganda, outperforming existing alternatives.
• The method exhibits a graceful degradation in performance with decreasing training data size, maintaining robustness even when limited labeled scans are available.
• Computational complexity is significantly lower than DDLS and patch-based SRC, scaling as O(NK) with number of training samples N and dictionary size K.
• Memory footprint is substantially reduced—by a factor of Ix×Iy compared to DDLS—due to shared dictionaries across image pixels rather than per-pixel dictionaries.
• The method achieves accurate separation of subdural hematomas from brain and CSF, even when subdural intensities overlap with those of CSF or brain tissue.
• The framework enables reliable volume estimation of brain and CSF, which is critical for evaluating neurocognitive outcomes after shunt surgery.