Efficient registration of pathological images: A joint PCA/image-reconstruction approach

Xu Han, Xiao Yang, Stephen R. Aylward, Roland Kwitt, Marc Niethammer

Introduction

Medical image registration—the process of aligning a patient's scan to a standard anatomical atlas—is a cornerstone of modern diagnosis and treatment planning. However, this process faces a significant hurdle when pathologies such as tumors, strokes, or traumatic brain injuries are present. These abnormalities create a "correspondence problem": because a standard healthy atlas lacks these features, traditional registration algorithms often attempt to warp healthy tissue to fit the tumor, leading to substantial anatomical errors.

The Limitations of Current Approaches

Existing solutions, notably Low-Rank/Sparse (LRS) decomposition, attempt to separate images into a "normal" (low-rank) component and a "pathology" (sparse) component. While theoretically sound, LRS suffers from three major practical drawbacks:

  1. Computational Intensity: It requires slow, iterative Singular Value Decompositions (SVD) across the entire dataset.
  2. Memory Constraints: The need to hold a full population of images in memory makes 3D processing nearly impossible on standard hardware.
  3. Image Blurring: LRS often blurs healthy tissue, which can paradoxically degrade registration accuracy in non-pathological regions.

Proposed Method: Joint PCA and Image Reconstruction

This paper introduces a fast, memory-efficient alternative that utilizes Principal Component Analysis (PCA) to "strip away" pathologies, creating a "quasi-normal" version of the image for accurate registration without requiring manual tumor masks.

The Two-Step Process

  1. Pre-computing a "Normal" Basis: The researchers first register a population of healthy images (from the OASIS dataset) to an atlas. PCA is then used to build a mathematical basis representing "normal" anatomy in atlas space. This step is performed only once.
  2. Quasi-Normal Reconstruction: When a new pathological image is processed, it is decomposed into two parts:
  3. The Quasi-Normal Part: A version of the image reconstructed to fit within the pre-computed PCA "normal" space.
  4. The Abnormal Part: Identified using Total Variation (TV) regularization. Unlike simple sparsity, TV is designed to capture spatially contiguous "blobs," making it far more effective at isolating solid tumors.
Decomposition of a pathological brain image into quasi-normal (top) and abnormal (bottom) components. The proposed method isolates the pathology while preserving sharp details in healthy tissue.

By decoupling the learning of normal anatomy from the processing of individual patient images, the algorithm significantly reduces computational overhead.

Experimental Results

The model was evaluated using the OASIS dataset for normal anatomy and the BRATS 2015 dataset for realistic tumor data. The results demonstrate that the PCA-based approach outperforms standard LRS across all key metrics:

  • Efficiency: In 3D applications, the PCA approach reduced processing time from over 6 hours to approximately 40 minutes per case. Memory requirements dropped from over 40 GB to a manageable 4–8 GB.
  • Accuracy: Registration error in tumor regions was significantly lower than direct registration. Critically, the method maintained high accuracy in healthy cortical tissue where LRS typically fails due to blurring.
  • Automation: The system achieved performance levels comparable to "cost function masking"—the current gold standard—but without the need for time-consuming manual tumor segmentation.
Quantitative comparison of registration errors across different brain regions. The proposed methods (represented among categories A-F) show improved stability and lower error rates in both pathological and healthy areas.

Conclusion

The Joint PCA/Image-Reconstruction approach offers a scalable solution for registering pathological images. By leveraging a pre-computed model of normal anatomy and employing Total Variation to isolate contiguous abnormalities, the method provides a high-speed, low-memory alternative to traditional decomposition techniques. Future work will focus on validating this approach across more diverse 3D datasets and different types of pathologies, such as strokes and lesions.

Summary updated

Abstract

Registration involving one or more images containing pathologies is challenging, as standard image similarity measures and spatial transforms cannot account for common changes due to pathologies. Low-rank/Sparse (LRS) decomposition removes pathologies prior to registration; however, LRS is memory-demanding and slow, which limits its use on larger data sets. Additionally, LRS blurs normal tissue regions, which may degrade registration performance. This paper proposes an efficient alternative to LRS: (1) normal tissue appearance is captured by principal component analysis (PCA) and (2) blurring is avoided by an integrated model for pathology removal and image reconstruction. Results on synthetic and BRATS 2015 data demonstrate its utility.
BibTeX
@inproceedings{han2017efficient,
  author = {Xu Han and Xiao Yang and Stephen R. Aylward and Roland Kwitt and Marc Niethammer},
  title = {{Efficient registration of pathological images: A joint PCA/image-reconstruction approach}},
  booktitle = {ISBI},
  year = {2017},
  url = {https://openreview.net/forum?id=HBxNjmE01x5},
}
Figure