CINDERELLA: Computerised INference and DERivation of MRI sequence Linguistic Labelling Algorithm
Poster No:
1914
Submission Type:
Abstract Submission
Authors:
Ivar Wamelink1, Joost Kuijer1, Vera Keil1, Frederik Barkhof2, Alle Meije Wink2
Institutions:
1Amsterdam University Medical Centre, Amsterdam, Netherlands, 2Amsterdam University Medical Centre, Amsterdam, Noord-Holland
First Author:
Co-Author(s):
Introduction:
Identifying the scanning protocol of MR DICOM datasets by their series descriptions is problematic for two reasons: names for the same sequence may vary between vendors and the series description is a free-form text field prone to inconsistencies.
Recent attempts at automatic protocol detection rely on machine learning of image intensities or DICOM header fields. [van der Voort et al., 2021; Bartnik et al., 2023]. This project aims to extract the protocol by exclusively comparing a minimal set of DICOM fields representing common acquisition parameters.
Our motivations are that relying on pixel intensities is (i) computationally costly, (ii) will misclassify faulty scans, and (iii): non-acquisition parameters may vary between vendors, institutes, field strengths and software versions.
We describe the implementation of our system and demonstrate its performance in a multi-site, multi-vendor, multi-sequencs imaging study.
Recent attempts at automatic protocol detection rely on machine learning of image intensities or DICOM header fields. [van der Voort et al., 2021; Bartnik et al., 2023]. This project aims to extract the protocol by exclusively comparing a minimal set of DICOM fields representing common acquisition parameters.
Our motivations are that relying on pixel intensities is (i) computationally costly, (ii) will misclassify faulty scans, and (iii): non-acquisition parameters may vary between vendors, institutes, field strengths and software versions.
We describe the implementation of our system and demonstrate its performance in a multi-site, multi-vendor, multi-sequencs imaging study.
Methods:
Our in-house database contains 206,769 series of 12,712 different scan sessions of 1,384 subjects from scanners of 4 vendors (162 different software versions among all vendors) and with different field strengths (GE: 1.5T and 3T; Siemens: 1T, 1.5T, 3T; Philips: 0.5T, 1T, 1.5T, 3T; Toshiba: 1.5T, 3T). Seven common acquisition-specific DICOM fields were extracted: FieldStrength, EchoTime, FlipAngle, RepetitionTime, InversionTime, DiffusionBValue, and Contrast.
We used a manual iterative process guided by expertise to create a vendor and field strength-specific database that could automatically label 16 sequence types (figure 1): 3D FLAIR, 2D FLAIR, 3D T1, 3D T1c, 2D T2, 2D dual-echo T2, DWI, 2D T1, 2D T1c, DSC, SWI, B0, B1, ASL, APT-CEST, and fMRI. Parameters were either matched to values from a set (e.g., vendors) or within a numerical range (e.g., echo time). The selected sequence was the one meeting all parameter criteria. Figure 1 illustrates the iterative process of creating the label database.
The protocol identification script was run on our in-house data set and did not classify derived series such as MPR. Correct identification was based on expert comparison of the predicted standardised label with the series description.
We used a manual iterative process guided by expertise to create a vendor and field strength-specific database that could automatically label 16 sequence types (figure 1): 3D FLAIR, 2D FLAIR, 3D T1, 3D T1c, 2D T2, 2D dual-echo T2, DWI, 2D T1, 2D T1c, DSC, SWI, B0, B1, ASL, APT-CEST, and fMRI. Parameters were either matched to values from a set (e.g., vendors) or within a numerical range (e.g., echo time). The selected sequence was the one meeting all parameter criteria. Figure 1 illustrates the iterative process of creating the label database.
The protocol identification script was run on our in-house data set and did not classify derived series such as MPR. Correct identification was based on expert comparison of the predicted standardised label with the series description.
·Figure 1. Flowchart that shows the iterative steps for creating the acquisition parameter database fo 16 sequence types.
Results:
For our in-house imaging data set, the outcomes were:
- Of the total of 206,769 series, 97,748 series were derivatives, e.g. multiplanar reformation (MPR)
- Out of the remaining 109,021 series, 89,462 were classified
- the remaining 19,559 did not meet the 7 acquisition parameter criteria
The majority of this latter group were preparation sequences such as scout (localiser) or calibration scans (12,160), as well as sequences that have not yet been added to the database, such as 3D T2-weighted scans and time of flight angiography.
In some cases, the classification by our algorithm did not correspond to the series description. These series turned out to be slices from different acquisitions that had erroneously got the same series identifier.
Based on the series description, the algorithm was 99.69%, 99.99% and 94.43% accurate for contrast-enhanced 3D T1-weighted scans, 3D FLAIR, and T2-weighted scans, respectively.
- Of the total of 206,769 series, 97,748 series were derivatives, e.g. multiplanar reformation (MPR)
- Out of the remaining 109,021 series, 89,462 were classified
- the remaining 19,559 did not meet the 7 acquisition parameter criteria
The majority of this latter group were preparation sequences such as scout (localiser) or calibration scans (12,160), as well as sequences that have not yet been added to the database, such as 3D T2-weighted scans and time of flight angiography.
In some cases, the classification by our algorithm did not correspond to the series description. These series turned out to be slices from different acquisitions that had erroneously got the same series identifier.
Based on the series description, the algorithm was 99.69%, 99.99% and 94.43% accurate for contrast-enhanced 3D T1-weighted scans, 3D FLAIR, and T2-weighted scans, respectively.
·Figure 2. Partial table showing predictions for 6 of the 16 sequence types, for all the vendors and field strengths appearing in our in-house imaging study.
Conclusions:
We have developed CINDERILLA to label MRI scans from a DICOM data set based upon a limited number of acquisition-related parameters, to correctly label 16 sequences. Focusing on acquisition parameters directly related to image contrast means that, unlike pixel-based protocol prediction, our method is not affected by faulty scans.
For performance evaluation, a human expert compared the predicted protocol to the series description. The prediction performance of T2-weighted scans was lower than others, because of overlapping sequence parameters between T2-weighted and diffusion-weighted scans, and requires further development.
We will run our software, developed in Python3.9, on further imaging study datasets, to assess prediction performance on new data that was not used to set the parameter ranges. We will make the software available after these additional tests.
For performance evaluation, a human expert compared the predicted protocol to the series description. The prediction performance of T2-weighted scans was lower than others, because of overlapping sequence parameters between T2-weighted and diffusion-weighted scans, and requires further development.
We will run our software, developed in Python3.9, on further imaging study datasets, to assess prediction performance on new data that was not used to set the parameter ranges. We will make the software available after these additional tests.
Modeling and Analysis Methods:
Methods Development 1
Neuroinformatics and Data Sharing:
Databasing and Data Sharing
Workflows 2
Informatics Other
Keywords:
Data analysis
Data Organization
MRI
Open-Source Code
Workflows
1|2Indicates the priority used for review
Provide references using author date format
van der Voort SR, Smits M, Klein S, for the Alzheimer’s Disease Neuroimaging Initiative (2021): DeepDicomSort: An Automatic Sorting Algorithm for Brain Magnetic Resonance Imaging Data. Neuroinformatics 19:159–184.