A 3D Multi-task Regression and Ordinal Regression Deep Neural Network for Collateral Imaging from Dynamic Susceptibility Contrast-Enhanced MR perfusion in Acute Ischemic Stroke

https://doi.org/10.1016/j.cmpb.2022.107071Get rights and content

Highlights

  • We propose a multi-task 3D deep learning approach for collateral imaging that take advantages of both regression and ordinal regression.

  • We introduce a novel discretization strategy, network architecture, and loss functions that are specifically designed for the generation of collateral maps.

  • We achieve the state-of-the-art performance on a dataset with >800 subjects as compared to 8 other competing methods.

Abstract

Background and Objective

Cerebral collaterals have been identified as one of the primary determinants for treatment options in acute ischemic stroke. Several works have been proposed, but these have not been adopted for a routine clinical usage due to their manual and heuristic nature as well as inconsistency and instability of the assessment. Herein, we present an advanced deep learning-based method that can automatically generate a multiphase collateral imaging (collateral map) derived from dynamic susceptibility contrast-enhanced MR perfusion (DSC-MRP) in an accurate and robust manner.

Methods

We develop a 3D multi-task regression and ordinal regression deep neural network for generating collateral maps from DSC-MRP, which formulates the prediction of collateral maps as both a regression task and an ordinal regression task. For an ordinal regression task, we introduce a spacing-decreasing discretization (SDD) strategy to represent the intensity of the collateral status on a discrete, ordinal scale. We also devise loss functions to achieve effective and efficient multi-task learning.

Results

We systematically evaluated the performance of the proposed network using DSC-MRP from 802 patients. On average, the proposed network achieved ≥0.900 squared correlation coefficient (R-Squared), ≥0.916 Tanimoto measure (TM), ≥0.0913 structural similarity index measure (SSIM), and ≤0.564 × 10−1 mean absolute error (MAE), outperforming eight competing models that have been recently developed in medical imaging and computer vision. We also found that the proposed network could provide an improved contrast between the low and high intensity regions in the collateral maps, which is a key to an accurate evaluation of the collateral status.

Conclusions

The experimental results demonstrate that the proposed network is able to generate collateral maps with high accuracy, facilitating a timely and prompt assessment of the collateral status in clinlcs. The future study will entail the optimization of the proposed network and its clinical evalution in a prospective manner.

Introduction

Endovascular recanalization and reperfusion have recently been adopted as a major treatment option for patients with acute ischemic stroke due to large-vessel occlusion. Although the treatment can reestablish antegrade perfusion to the ischemic brain, the treatment has a disproportionate effect on the patients’ outcome. Prior studies have shown that the outcome is largely dependent on the cerebral collaterals, i.e., an alternative, auxiliary vascular pathway that maintains the cerebral blood flow when the primary source of blood flow has been compromised by the occlusion. For instance, patients with a large ischemic core and poor collateral perfusion status are strongly associated with hemorrhagic complications and poor recanalization rates and functional outcomes [1,2]. A better collateral status is highly related to less infarct growth, better functional outcomes, and as extension of the time window for endovascular treatment [3], [4], [5]. Therefore, an accurate and reliable method that can assess the collateral status is of great importance in determining the treatment option for the patients with acute ischemic stroke.

A number of imaging-based methods have been proposed to evaluate the collateral status. Both single- and multi-phase computed tomography angiography (CTA) have been employed to score the arterial status and to determine the patients’ outcome [6]. Single-phase CTA tends to underestimate the leptomeningeal collaterals with a longer transit time owing to early triggering of CT scans. Multi-phase CTA improves upon the estimation of the leptomeningeal collaterals but still retains the issues with radiation exposure, iodine contrast media, and inaccurate infarct volume estimation. Moreover, collateral maps, derived from dynamic susceptibility contrast-enhanced MR perfusion (DSC-MRP), have been proposed to evaluate the patients with acute ischemic stroke [7]. Some studies demonstrate that the collateral status can serve as an indicator of final infarct size and functional outcomes for the patients who are eligible for intra-arterial thrombectomy and intravenous thrombolysis [1,8]. In line with these developments, another approach of deriving and evaluating collateral maps has been developed from dynamic contrast-enhanced MR angiography [9,10]. The collateral maps derived from DSC-MRP and MR angiography are almost identical to each other [9]. The grading system for the collateral maps shows a significant correlation with the patients’ outcome [10]. Even though these methods have shown to be effective in assessing the collateral status, they usually rely on heuristic assumptions/decisions during the assessment of the collateral status and/or require several manual processing steps, potentially leading to inconsistency and instability in the assessment and decision making for acute ischemic stroke. Alternatively, an artificial intelligence system that can automatically and objectively measure the collateral status could facilitate improved evaluation and prediction of patients’ outcome.

Herein, we propose a 3D multi-task regression and ordinal regression deep neural network, designated as 3D-MROD-Net, for collateral imaging from DSC-MRP (Fig. 1). 3D-MROD-Net simultaneously performs a regression task and an ordinal regression task. The regression task aims to predict a continuous value, i.e., the intensity of the collaterals. The goal of the ordinal regression task is to predict a discrete value (or class label) on an ordinal scale. The class label represents the relative rank of the intensity of the collaterals. The proposed network adopts pyramid pooling and utilizes two branches: one for the regression task and the other for the ordinal regression task. For effective and efficient multi-task learning, we devise loss functions and introduce a spacing-decreasing discretization (SDD) strategy. To systematically evaluate the performance of the proposed network, a number of DSC-MRP images from stroke and healthy subjects are employed. The experimental results demonstrate that the proposed network achieves the state-of-the-art performance in the generation of collateral maps as compared to other competing methods. Fig. 2 demonstrates the workflow of the proposed method.

The main contributions of the proposed approach can be summarized as follows:

  • We propose a multi-task 3D deep learning approach for collateral imaging that take advantages of both regression and ordinal regression.

  • We develop a novel discretization strategy, i.e., SDD strategy, to produce discrete collateral maps that can address imbalance in the ground truth.

  • We propose a network architecture that exploits the recent developments in the architecture and is tailored to the generation of collateral maps.

  • We introduce loss functions with task-specific uncertainty for an efficient learning of both regression and ordinal regression tasks.

  • We achieve the state-of-the-art performance on a dataset with >800 subjects as compared to 8 other competing methods.

Section snippets

Deep learning for stroke

Recently, artificial intelligence, in particular deep learning [11], has revolutionized data/image processing and analysis as well as data interpretation and decision making. Among various deep learning algorithms, a deep convolution neural network (CNN) has gained popularity as an effective approach in various medical image problems, namely classification, object detection and segmentation, image registration, image generation and enhancement [12], [13], [14]. This approach has been

Acute ischemic stroke dataset

This study contains 802 consecutive subjects including 344 healthy subjects (controls) and 458 acute ischemic stroke patients (stroke cases) recruited from two university medical centers from November 2015 to March 2020. The normal and stroke cases are separately divided into training dataset, validation dataset, and test dataset, including 401, 172, and 229 cases, respectively. Specifically, the training dataset contains 172 normal cases from January 2016 to September 2019 and 229 stroke cases

Results

Table 2 and Fig. 4 demonstrate the performance of 3D-MROD-Net and other competing models on the test dataset of 104 control and 136 stroke cases. For all five phases (Art, Cap, Even, LVen, and Del), 3D-MROD-Net was able to generate collateral maps in an accurate and robust manner, achieving ≥0.900 R-Squared (Eq. (6)), ≥0.916 TM (Eq. (7)), ≥0.0913 SSIM (Eq. (8)), and ≤0.564 MAE (x10−1) (Eq. (9)) on average. Among the five phases, the 3D-MROD-Net generally obtained lower performance in Art than

Discussion

Assessing the collateral status plays a vital role in a decision-making process for the treatment of stroke patients. Several methods have been proposed but rarely adopted for a routine clinical service. This is mainly due to manual processing and heuristic decisions, potentially leading to an uncertainty in the assessment of the collateral status. In this study, we propose an automated deep learning model that can generate five-phase collateral maps that are essential for an accurate

Conclusions

In this study, we present 3D-MROD-Net for collateral imaging in acute ischemic stroke. Equipped with an advanced deep learning and multi-task learning, 3D-MROD-Net was able to generate collateral imaging for both stroke cases and controls with high accuracy. The ability to automatically generate collateral imaging could facilitate a timely and prompt assessment of the collateral status, which is critical to acute ischemic stoke. The future study will entail the optimization of 3D-MROD-Net for

Conflict of Interest

The authors declare that they have no conflict of interest.

Acknowledgement

This study was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea governemnt (No. NRF-2021R1A2C2014557, No. NRF-2021K1A3A1A88100920, No. NRF-2021R1A4A1031864, and No. NRF-2020R1F1A1071619).

References (51)

  • S.J. Kim

    A novel magnetic resonance imaging approach to collateral flow imaging in ischemic stroke

    Ann. Neurol.

    (2014)
  • P. Seners

    Better collaterals are independently associated with post-thrombolysis recanalization before thrombectomy

    Stroke

    (2019)
  • H.G. Roh

    A novel collateral imaging method derived from time-resolved dynamic contrast-enhanced MR angiography in acute ischemic stroke: A Pilot Study

    AJNR Am. J. Neuroradiol.

    (2019)
  • H.J. Kim

    Multiphase MR angiography collateral map: functional outcome after acute anterior circulation ischemic stroke

    Radiology

    (2020)
  • Y. LeCun et al.

    Deep learning

    Nature

    (2015)
  • B.K. Menon

    Multiphase CT Angiography: A New tool for the imaging triage of patients with acute ischemic stroke

    Radiology

    (2015)
  • A. Nielsen

    Prediction of tissue outcome and assessment of treatment effect in acute ischemic stroke using deep learning

    Stroke

    (2018)
  • C. Lucas

    Learning to predict ischemic stroke growth on acute CT perfusion data by interpolating low-dimensional shape representations

    Front. Neurol.

    (2018)
  • N. Stier

    Deep learning of tissue fate features in acute ischemic stroke

  • A. Pinto

    Stroke lesion outcome prediction based on MRI imaging combined with clinical information

    Front Neurol

    (2018)
  • S.M. Cui et al.

    Brain tumor automatic segmentation using fully convolutional networks

    J. Med. Imag. Health Inf.

    (2017)
  • D. Robben et al.

    Perfusion parameter estimation using neural networks and data augmentation

  • A. Hess

    Synthetic Perfusion Maps: Imaging Perfusion Deficits in DSC-MRI with Deep Learning

    (2019)
  • M.N.N. To

    Deep regression neural networks for collateral imaging from dynamic susceptibility contrast-enhanced magnetic resonance perfusion in acute ischemic stroke

    Int. J. Comput. Assist. Radiol. Surg.

    (2020)
  • Y. Zhang et al.

    arXiv preprint

    (2018)
  • Cited by (0)

    View full text