Click here to

  1. Home
  2. Meeting Program
  3. All Sessions
  4. Abstracts
Session: MRI Radiomics and Segmentation [Return to Session]

Weakly Supervised Convolutional Neural Networks for Segmentation of Diffusely Abnormal White Matter in Multiple Sclerosis

B Musall1*, A Kamali1, J Lincoln1, V Ly2, X Luo2, P Narayana1, R Gabr1, (1) University of Texas McGovern Medical School, Houston, TX, (2) University of Texas School of Public Health, Houston, TX

Presentations

TU-D930-IePD-F9-2 (Tuesday, 7/12/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 9

Purpose: Multiple sclerosis (MS) lesion load, as assessed on T2w MRI, is only modestly correlated with clinical disease symptoms. Diffusely Abnormal White Matter (DAWM) is frequently present in MS and is being investigated as an MS biomarker. However, segmentation of DAWM is very difficult and time consuming. In this work, we trained U-Nets with weak supervision to address the challenge of DAWM segmentation. The trained U-Nets were then refined using a small number of cases with expert manual segmentation.

Methods: This study used 617 baseline MRI datasets of relapse-remitting MS patients from the CombiRx trial (NCT00211887). MR data included FLAIR, T2w, pre-contrast T1w, and proton density-weighted series. Two expert readers segmented DAWM on a randomly-selected subset of 20 patients.Two heuristic methods were applied to segment DAWM on the MRI. U-Nets were trained to convergence for the same task under weak supervision by using the imperfect ground truth generated by the heuristic methods. UNets were then refined to convergence at a lower learning rate using 10 of the reader-segmented cases. For both training steps, 20% of the dataset was used for validation to confirm convergence of the UNets.In the remaining 10 reader-segmented cases, DAWM segmentations were compared on a patient-wise basis using Dice Similarity Coefficient (DSC) to assess spatial agreement and using Spearman’s rho to assess correlation of segmented DAWM volumes.

Results: Inter-reader comparison showed an average DSC of 0.49±0.07 and a correlation of 0.95. The weakly-supervised UNet DSC performance (reader 1/reader 2) showed DSCs of 0.23±0.08/0.24±0.05 and correlations of 0.85/0.84. The refined UNet showed improved DSCs of 0.34±0.17/0.32±0.14 and slightly improved correlations of 0.87/0.89.

Conclusion: Weakly-supervised U-Nets refined with expert segmentation show promise for DAWM assessment in MS patients.

Keywords

Segmentation, MRI, Brain

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email

Share:

Additional Links