[Paper Review] Towards Clinical Diagnosis: Automated Stroke Lesion Segmentation on Multimodal MR Image Using Convolutional Neural Network

Zhiyang Liu, Chen Cao+4 more|arXiv (Cornell University)|Mar 5, 2018

Brain Tumor Detection and Classification16 references3 citations

TL;DR

This paper proposes a residual-structured fully convolutional network (Res-FCN) with atrous convolution and global convolution networks to automatically segment ischemic stroke lesions from multimodal MR images (DWI, ADC, T2WI). The method achieves a mean Dice coefficient of 0.645 and only 1.515 false negative lesions per patient—approaching radiologist-level performance—making it highly suitable for clinical deployment in stroke diagnosis.

ABSTRACT

The patient with ischemic stroke can benefit most from the earliest possible definitive diagnosis. While the high quality medical resources are quite scarce across the globe, an automated diagnostic tool is expected in analyzing the magnetic resonance (MR) images to provide reference in clinical diagnosis. In this paper, we propose a deep learning method to automatically segment ischemic stroke lesions from multi-modal MR images. By using atrous convolution and global convolution network, our proposed residual-structured fully convolutional network (Res-FCN) is able to capture features from large receptive fields. The network architecture is validated on a large dataset of 212 clinically acquired multi-modal MR images, which is shown to achieve a mean dice coefficient of 0.645 with a mean number of false negative lesions of 1.515. The false negatives can reach a value that close to a common medical image doctor, making it exceptive for a real clinical application.

Motivation & Objective

• To develop an automated, accurate, and clinically viable method for ischemic stroke lesion segmentation in multimodal MRI.
• To address the challenge of limited training data and high variability in lesion appearance across patients and scanners.
• To reduce false negative detections, which are critical in avoiding misdiagnosis in acute stroke.
• To leverage multimodal MRI (DWI, ADC, T2WI) for improved lesion detection and artifact reduction.
• To achieve performance comparable to expert radiologists in lesion segmentation, particularly in minimizing false negatives.

Proposed method

• The proposed Res-FCN uses residual learning to train a very deep fully convolutional network, improving gradient flow and training stability.
• Atrous convolution is employed to expand the receptive field without increasing parameters or reducing resolution.
• Global convolution networks (GCN) are integrated to capture long-range contextual dependencies across the image.
• The network is trained end-to-end on a dataset of 212 clinically acquired multimodal MR images (DWI, ADC, T2WI).
• A threshold-optimization strategy is applied to balance false positives and false negatives, with performance evaluated using Dice coefficient and false positive/negative counts.
• The method is compared against EDD-Net and U-Net using standard metrics on a test set of 36 patients.

Experimental results

Research questions

• RQ1 Can a deep learning-based segmentation model achieve high accuracy and low false negative rates in ischemic stroke lesion detection on multimodal MRI?
• RQ2 How does the integration of atrous and global convolutional layers improve lesion segmentation in low-data regimes?
• RQ3 To what extent does using multimodal MRI (DWI, ADC, T2WI) reduce artifacts and improve detection compared to single-modal input?
• RQ4 How does the proposed Res-FCN compare to state-of-the-art models (EDD-Net, U-Net) in terms of false positive and false negative trade-offs?
• RQ5 Can the model’s performance approach that of a clinical radiologist in terms of false negative detection?

Key findings

• The proposed Res-FCN achieved a mean Dice coefficient of 0.645 on the test dataset, outperforming both EDD-Net (0.644) and U-Net (0.557).
• The model recorded only 1.515 false negative lesions per patient on average, approaching the performance of a typical clinical radiologist.
• With a threshold of δ=1, the Res-FCN achieved a Dice coefficient of 0.658 while maintaining only 2.485 false positives and 1.866 false negatives per patient.
• The Res-FCN demonstrated superior FN-FP trade-off, reducing false positives by approximately 2 compared to EDD-Net at similar false negative levels.
• The use of multimodal MRI (DWI, ADC, T2WI) significantly improved lesion detection accuracy and reduced artifacts compared to single-sequence methods.
• The model’s performance was robust across diverse patient scans, indicating strong generalization despite limited training data.