MRI Detection of Neonatal Hypoxic Ischemic Encephalopathy: Machine v.s. Radiologists
Stand-By Time
Monday, June 26, 2017: 12:45 PM - 2:45 PMSubmission No:
1662
Submission Type:
Abstract Submission
On Display:
Monday, June 26 & Tuesday, June 27
Authors:
Yangming Ou1, Randy Gollub2, Jing Wang3, Qianqian Fan3, Sara Bates4, Joseph Chou4, Rebecca Weiss4, Kallirroi Retzepis4, Steve Pieper5, Camilo Jaimes4, Shawn Murphy4, Lilla Zöllei6, P. Ellen Grant7
Institutions:
1Boston Children's Hospital, Boston, United States, 2MGH, Charlestown, MA, 3Union Hospital, Tongji Medical College, Wuhan, Hubei, 4MGH, Boston, MA, 5Isomics Inc., Boston, MA, 6Athinoula A Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA, 7Boston Children's Hospital, Harvard Medical School, Boston, MA
First Author:
Introduction:
Hypoxic ischemic encephalopathy (HIE) is a brain injury that affects 1-6/1000 neonates in the first 2 weeks of life [1]. Radiologists interpret HIE from apparent diffusion coefficient (ADC) maps. Excessively low ADC values indicating restricted water flow warrant the reads of HIE lesions [2]. This interpretation has uncertainties [3]. Uncertainties mainly come from the lack of knowledge about normal variation of ADC values among healthy neonates. Given the lack of knowledge about normal ADC variation, it remains a qualitative and subjective judgment to when and where an ADC value is excessively low [3]. We recently constructed first of its kind normative ADC atlases [4,5] (released at https://www.nitrc.org/projects/mgh_adcatlases). Our atlases quantify normal ADC variations at each brain voxel location in a normative neonatal population. This gave rise to the possibility of machine-assisted, quantitative and objective interpretation of HIE [6]. Herein, we quantitatively compare atlas-based machine interpretation of HIE lesions with the interpretations from two practicing radiologists, in order to test the diagnostic utility of our atlases.
Methods:
Our constructed normative ADC atlases represent an average anatomy/geometry in a normative neonate cohort [5]. At each voxel, our atlases quantified the mean (μ) and standard deviation (σ) ADC values (Figure 1, left panel) [5].
Our evaluation dataset contains ADC maps of 8 HIE neonates, diagnosed and treated in Boston Children's Hospital, scanned on a Siemens 3T scanner with b=1000. Voxel size was 2x2x2mm and each ADC map contained 128x128x58 voxels.
We non-rigidly registered the atlases into the neonatal patient's space [7] to enable voxel-by-voxel comparison between the patient and the normative atlas. After the registration, patient's ADC value Ix at voxel x was converted to a Z value (Zx=(Ix - μx)/σx), which quantified how many standard deviations away this patient's ADC value was from the mean normal ADC value at the corresponding anatomical location. We labeled those voxels as being affected by HIE lesions if they had Z values lower than a threshold t. We measured the accuracy of atlas-based automatic HIE detection as the threshold t varied. The accuracy was defined as the Dice overlap coefficient with manual annotations of HIE-affected regions from the two radiologists (Rater1 and Rater2, each with 7+ years of experience). We report Rater1-vs-Rater2 Dice overlap, and the average of Rater1-vs-Algorithm and Rater2-vs-Algorithm Dice overlap as the algorithm's threshold t varied.
Our evaluation dataset contains ADC maps of 8 HIE neonates, diagnosed and treated in Boston Children's Hospital, scanned on a Siemens 3T scanner with b=1000. Voxel size was 2x2x2mm and each ADC map contained 128x128x58 voxels.
We non-rigidly registered the atlases into the neonatal patient's space [7] to enable voxel-by-voxel comparison between the patient and the normative atlas. After the registration, patient's ADC value Ix at voxel x was converted to a Z value (Zx=(Ix - μx)/σx), which quantified how many standard deviations away this patient's ADC value was from the mean normal ADC value at the corresponding anatomical location. We labeled those voxels as being affected by HIE lesions if they had Z values lower than a threshold t. We measured the accuracy of atlas-based automatic HIE detection as the threshold t varied. The accuracy was defined as the Dice overlap coefficient with manual annotations of HIE-affected regions from the two radiologists (Rater1 and Rater2, each with 7+ years of experience). We report Rater1-vs-Rater2 Dice overlap, and the average of Rater1-vs-Algorithm and Rater2-vs-Algorithm Dice overlap as the algorithm's threshold t varied.
Results:
The two radiologists had Dice overlap over 0.5 in 6 out of 8 patients (min 0.59, max 0.89, median 0.82). Figure 1 shows two radiologists' reads of HIE lesions (red and green), and automated results based on thresholding the Z-map at various levels. We further quantitatively compared the inter-rater Dice overlap and algorithm-rater Dice overlaps at various thresholds, for each of the 6 patients (Figure 2) and for all patients combined (Figure 3). As Figure 3 shows, threshold value -1.2 gave an algorithm-rater agreement that was closest to inter-rater agreement.
·Figure 1. HIE detection by two radiologists (Rater 1 and 2) and by atlas-based Z map calculation and thresholding.
·Figure 2. Inter-rater Dice overlap as compared to algorithm-rater Dice overlap at various threshold values, for each of the 6 patients. 3: Summarizing sub-figures in Figure 2 into one unified figure,
Conclusions:
Our constructed normative neonatal ADC atlases provided a standard reference for automated, quantitative and objective detection of HIE lesions. The algorithm agreed to radiologists at a level comparable to radiologists' agreement. Future work includes (a) allowing the threshold to automatically adapt to different patients and different brain regions; (b) developing more sophisticated machine learning algorithms using the current results as initializations; and (c) further validating algorithm's accuracy by comparing to annotations of more patients from more radiologists (especially experienced pediatric/neonatal neuro-radiologists).
Disorders of the Nervous System:
Stroke
Informatics:
Brain Atlases 1
Informatics Other
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 2
Segmentation and Parcellation
Keywords:
Atlasing
Data analysis
Informatics
Pediatric Disorders
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Would you accept an oral presentation if your abstract is selected for an oral session?
I would be willing to discuss my abstract with members of the press should my abstract be marked newsworthy:
Please indicate below if your study was a "resting state" or "task-activation” study.
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute the presentation in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels or other electronic media and on the OHBM website.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Internal Review Board (IRB) or Animal Use and Care Committee (AUCC) Approval. Please indicate approval below. Please note: Failure to have IRB or AUCC approval, if applicable will lead to automatic rejection of abstract.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references in author date format
2. Wolf, R. L., Zimmerman, R. A., Clancy, R. & Haselgrove, J. H. Quantitative Apparent Diffusion Coefficient Measurements in Term Neonates for Early Detection of Hypoxic-Ischemic Brain Injury: Initial Experience 1. Radiology 218, 825–833 (2001).
3. Cheong, J. L., Coleman, L., Hunt, R. W., Lee, K. J., Doyle, L. W., Inder, T. E., Jacobs, S. E. & Collaboration, I. C. E. Prognostic utility of magnetic resonance imaging in neonatal hypoxic-ischemic encephalopathy: substudy of a randomized trial. Arch. Pediatr. Adolesc. Med. 166, 634–640 (2012).
4. Ou, Y., Gollub, R. L., Retzepi, K., Reynold, N. A., Pienaar, R., Murphy, S. N., Grant, P. E. & Zöllei, L. Brain Extraction in Pediatric ADC Maps, toward Characterizing Neuro-Development in Multi-Platform and Multi-Institution Clinical Images. NeuroImage 122, 246–261 (2015).
5. Ou, Y., Reynolds, N., Gollub, R., Pienaar, R., Wang, Y., Wang, T., Sack, D., Andriole, K., Pieper, S. & Herrick, C. Developmental brain ADC atlas creation from clinical images. in Organization for Human Brain Mapping (OHBM) (2014).
6. Ou, Y., Jaims, C., Gollub, R. L., Retzepis, K., Bates, S., Murphy, S., Grant, P. E. & Zollei, L. Neonatal Brain Injury Detection in MRI: An Atlas-based Fully-Automatic Approach. in Human Brain Mapping Annual Meeting (2015).
7. Ou, Y., Sotiras, A., Paragios, N. & Davatzikos, C. DRAMMS: Deformable registration via attribute matching and mutual-saliency weighting. Med. Image Anal. 15, 622–639 (2011).