Sparse Correlated Diffusion Imaging: A New Computational Diffusion MRI Modality for Prostate Cancer Detection

Farzad Khalvati, Junjie Zhang, Sameer Baig, Masoom A. Haider, Alexander WongJournal of Computational Vision and Imaging Systems 2016 · DOI

Abstract

Diffusion weighted imaging (DWI) is a promising magnetic resonance imaging (MRI) modality with wide applications in the diagnosis of different types of diseases such as prostate cancer. DWI provides a large amount of imaging data which often makes it difficult to interpret accurately, mainly due to the fact that much of the information in diffusion imaging cannot be deciphered by human experts alone. Computational diffusion MRI (CD-MRI) aims to leverage computational means to generate imagery from diffusion signals which are easier to interpret by human experts. Recently, a new CD-MRI modality called correlated diffusion imaging (CDI) has been proposed which takes advantage of the joint correlation of diffusion signal attenuation across multiple gradient pulse strengths and timings to improve the separability of cancerous and healthy tissues. In this paper, we propose a new CD-MRI modality called Sparse CDI (sCDI) where an optimally sparse subset of diffusion signals contributes to the formation of the final diffusion signal, leading to further separation of cancerous and healthy tissue in the prostate gland compared to CDI and conventional DWI.

Introduction

Prostate cancer is a leading health concern for men, and early, accurate detection is critical for improving survival rates. Currently, clinical diagnosis relies heavily on multi-parametric MRI, specifically Diffusion-Weighted Imaging (DWI). DWI measures how water molecules move through tissue; because cancerous tumors are dense, they restrict this movement differently than healthy tissue.

However, DWI produces a massive amount of complex data that is notoriously difficult for humans to interpret. This leads to "inter-observer variability," where different radiologists might reach different conclusions from the same scan. While previous attempts like Correlated Diffusion Imaging (CDI) improved performance by looking at the joint correlation of signals across different pulse strengths, they treated all signals as equally important. The authors of this paper argue that "more data" isn't always better; noise from less relevant signals can actually obscure a tumor. Sparse CDI (sCDI) addresses this by selecting only the most informative signals.

Methodology

The core innovation of Sparse CDI (sCDI) is the introduction of a "sparse sampling vector" ($R$) into the signal mixing function.

  1. Signal Mixing and Sparsification: Traditional CDI combines multiple diffusion signals into a single image. sCDI adds a binary selection process to each signal. The team optimized the selection to find a specific subset of signals (using a "sparsification factor" of $m=3$) that maximizes diagnostic accuracy.
  2. Optimization: The selection was tuned to maximize the Area Under the ROC Curve (AUC), a standard metric for diagnostic performance.
  3. Synthetic High b-values: High "b-value" images, which show restricted diffusion more clearly, are often too noisy to acquire directly. The authors used a Bayesian model to computationally synthesize high b-value images ($1300$ and $2200\ s/mm^2$) from lower-strength acquired images ($0$ to $1000\ s/mm^2$).

Experimental Results

The study involved a cohort of 17 patients at the Sunnybrook Health Sciences Centre. The training set (9 patients) was used to optimize the algorithm, while the testing set (8 patients) with pathology-confirmed cancer was used for evaluation.

ROC curves for ADC, CDI, and sCDI showing the superior diagnostic performance of the sparse approach.

As shown in the ROC analysis, sCDI achieved an AUC of 0.92, a statistically significant improvement ($p < 0.05$) over the clinical standard (ADC) and the previous state-of-the-art (CDI).

Modality Area Under ROC (AUC)
ADC (Standard) 0.86
CDI 0.88
sCDI (Proposed) 0.92

Discussion

The "Less is More" Insight

The most striking insight is that discarding data can improve results. By forcing the algorithm to pick only three signals, the researchers filtered out the noise of the surrounding healthy tissue, making the cancerous regions stand out with much higher contrast.

Comparison of (a) ADC, (b) CDI, and (c) the proposed sparse modality (labeled eCDI in figure). The red arrows indicate the tumor location, which is most distinct in the sparse reconstruction.

Clinical Implications

By making the tumor visually "pop," sCDI reduces the burden on the radiologist's intuition, potentially leading to more consistent diagnoses. Furthermore, the paper demonstrates that better results can be achieved through computational synthesis rather than relying solely on more powerful (and expensive) MRI hardware.

Limitations and Future Work

The primary limitation is the small sample size of 17 patients. A larger, multi-center trial is needed to validate these results across different MRI manufacturers. Future work could involve dynamically determining the optimal number of signals for different tissue types or integrating sCDI into Computer-Aided Detection (CAD) systems to automatically flag suspicious regions for radiologists.

Summary updated

Abstract

Diffusion weighted imaging (DWI) is a promising magnetic resonanceimaging (MRI) modality with wide applications in diagnosisof different types of diseases such as prostate cancer. DWI providesa large amount of imaging data which often makes it difficultto interpret accurately, mainly due to the fact that much of informationin diffusion imaging cannot be deciphered by human expertsalone. Computational diffusion MRI (CD-MRI) aims to leveragecomputational means to generate imagery from diffusion signalswhich are easier to interpret by human experts. Recently, anew CD-MRI modality called correlated diffusion imaging (CDI) hasbeen proposed which takes advantage of the joint correlation of diffusionsignal attenuation across multiple gradient pulse strengthsand timings to improve the separability of cancerous and healthytissues. In this paper, we propose a new CD-MRI modality calledSparse CDI (sCDI) where an optimally sparse subset of diffusionsignals contributes to the formation of the final diffusion signal leadingto further separation of cancerous and healthy tissue in prostategland compared to CDI and conventional DWI.
BibTeX
@article{khalvati2016sparse,
  author = {Farzad Khalvati and Junjie Zhang and Sameer Baig and Masoom A. Haider and Alexander Wong},
  title = {{Sparse Correlated Diffusion Imaging: A New Computational Diffusion MRI Modality for Prostate Cancer Detection}},
  journal = {Journal of Computational Vision and Imaging Systems},
  year = {2016},
  doi = {10.15353/vsnl.v2i1.107},
}
Figure