[Paper Review] Multiple Sclerosis Lesion Segmentation from Brain MRI via Fully Convolutional Neural Networks

Snehashis Roy, John A. Butman+3 more|arXiv (Cornell University)|Mar 24, 2018

Brain Tumor Detection and Classification61 references62 citations

TL;DR

A fully convolutional CNN (FLEXCONN) with parallel multi-contrast pathways achieves MS lesion segmentation from multi-contrast MRI, outperforming several state-of-the-art methods on ISBI-61 and MS-100 datasets.

ABSTRACT

Multiple Sclerosis (MS) is an autoimmune disease that leads to lesions in the central nervous system. Magnetic resonance (MR) images provide sufficient imaging contrast to visualize and detect lesions, particularly those in the white matter. Quantitative measures based on various features of lesions have been shown to be useful in clinical trials for evaluating therapies. Therefore robust and accurate segmentation of white matter lesions from MR images can provide important information about the disease status and progression. In this paper, we propose a fully convolutional neural network (CNN) based method to segment white matter lesions from multi-contrast MR images. The proposed CNN based method contains two convolutional pathways. The first pathway consists of multiple parallel convolutional filter banks catering to multiple MR modalities. In the second pathway, the outputs of the first one are concatenated and another set of convolutional filters are applied. The output of this last pathway produces a membership function for lesions that may be thresholded to obtain a binary segmentation. The proposed method is evaluated on a dataset of 100 MS patients, as well as the ISBI 2015 challenge data consisting of 14 patients. The comparison is performed against four publicly available MS lesion segmentation methods. Significant improvement in segmentation quality over the competing methods is demonstrated on various metrics, such as Dice and false positive ratio. While evaluating on the ISBI 2015 challenge data, our method produces a score of 90.48, where a score of 90 is considered to be comparable to a human rater.

Motivation & Objective

• Demonstrate a fully convolutional CNN architecture for MS lesion segmentation from multi-contrast MR images.
• Assess the impact of using large 2D patches and convolution-based membership prediction on segmentation accuracy.
• Compare FLEXCONN to established MS lesion segmentation methods across public and private datasets.

Proposed method

• Use two parallel convolutional pathways for multi-contrast inputs (MPRAGE and FLAIR) to extract features.
• Concatenate pathway outputs and apply a second convolutional pathway to produce a lesion membership map.
• Train with patch-based supervision where training memberships are generated by Gaussian-blurred manual masks.
• Prefer fully convolutional prediction over a final fully connected layer to improve spatial coherence and reduce false positives.
• Optimize patch size (35x35) and pathway depth (five filter banks) using VAL-28 validation set.
• Threshold membership maps (0.30) to produce final binary segmentations; evaluate with Dice, LFPR, PPV, and VD.

Experimental results

Research questions

• RQ1 Can a fully convolutional CNN improve MS lesion segmentation accuracy over traditional methods?
• RQ2 What patch size and network depth yield the best balance between Dice score and false positives for MS lesion segmentation?
• RQ3 How does FLEXCONN compare to LesionTOADS, OASIS, LST, and S3DL on ISBI-61 and MS-100 datasets?

Key findings

• FLEXCONN achieves higher Dice and lower lesion false positive rate (LFPR) and volume difference (VD) than competing methods on MS-100.
• On MS-100, FLEXCONN median metrics: Dice 0.5639, LFPR 0.3077, PPV 0.6040, VD 0.1978.
• On ISBI-61, FLEXCONN reports the highest score among evaluated methods: Score 90.48, indicating segmentation accuracy comparable to human inter-rater reliability.
• Compared to FC-based predictions, the convolutional membership approach reduces false positives in cortex and yields faster whole-volume inference (seconds per 3D volume).
• Best performance with 35x35 patches and a 5-pathway depth; smaller patches underperform in Dice and LFPR.