RSNA 2016

Abstract Archives of the RSNA, 2016


SSC08-04

A Novel Bi-Input Convolutional Neural Network for Deconvolution-Free Estimation of Stroke MR Perfusion Parameters

Monday, Nov. 28 11:00AM - 11:10AM Room: S402AB



King Chung Ho, MSc, los angeles, CA (Presenter) Nothing to Disclose
Fabien Scalzo, PhD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Karthik V. Sarma, BSc, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Suzie M. El-Saden, MD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Alex A. Bui, PhD, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
Corey W. Arnold, Los Angeles, CA (Abstract Co-Author) Nothing to Disclose
PURPOSE

Perfusion magnetic resonance (MR) images are often used in conjunction with diffusion weighted images during the assessment of acute ischemic stroke to distinguish between the likely salvageable tissue and infarcted core. Methods such as singular value decomposition have been developed to approximate perfusion parameters from these images. However, studies have shown that existing deconvolution algorithms can introduce distortions that influence the measurements. In this work, we present a novel bi-input convolutional neural network (bi-CNN) to approximate perfusion parameters without deconvolution. We applied the trained bi-CNN to approximate cerebral blood volume (CBV).

METHOD AND MATERIALS

MR perfusion data was collected retrospectively for a set of 11 patients who had acute ischemic stroke. The ground truth perfusion maps (i.e., CBV) and arterial input functions (AIFs) were generated from ASIST-Japan perfusion mismatch analyzer, with the resulting CBV values ranging between 0-201 ml/100g. A set of 87,600 training patches with associated AIFs and CBVs were randomly sampled from the source perfusion data. Each patch had a size of 3 x 3 x 70 (width x height x time), and the center of the patch was the voxel of interest for estimation.Our bi-CNN is a 5-layer model with two parts: 1) two separate 3D convolutional and nonlinear layers for the training patch and its AIF, and 2) three fully-connected layers that combine the output of the first part to produce an estimated CBV. The model was trained with batch gradient descent, with a momentum of 0.9.

RESULTS

A leave-one-brain-out validation was performed to estimate voxel-wise CBV values. The bi-CNN achieved an average mean squared error (MSE) of 3.799 ml/100g +/-3.715. CBV deficits (< 2.5 ml/100g) could be identified from the bi-CNN estimated maps.

CONCLUSION

Our patch-based bi-CNN model is capable of estimating CBV in stroke patients. The model can be potentially extended to other disease domains, such as perfusion analysis in cancer. Future work includes experimenting on a larger dataset and estimating other important perfusion parameters, such as time-to-maximum (Tmax).

CLINICAL RELEVANCE/APPLICATION

Convolutional neural networks can be trained to approximate stroke MR perfusion parameters (e.g., CBV) and are a potential alternative method for automated quantification of perfusion abnormalities.