Deep learning for automatic bone marrow apparent diffusion coefficient measurements from whole-body MRI in patients with multiple myeloma – a retrospective multicenter study
- Supplemental Digital Content 1 -
Details on inclusion and exclusion process for different datasets
Development data set: In order to train a robust segmentation algorithm, we intended to use a data set which contains heterogenous data regarding both image acquisition parameters and disease stages (covering asymptomatic patients with few pathologies and patients with actual multiple myeloma with severe pathologies, including patients before and after treatment). Therefore, we chose to use a data set from a bi-institutional test-retest-study,1,2 in which data was acquired with 2 different 1.5 T MRI scanners and one 3T MRI scanners, as well as with 2 different DWI protocols at scanner 1, for the development set. Additionally, the patients represent a wide range of different stages of the disease, including all asymptomatic precursor stages, newly diagnosed multiple myeloma patients, and patients who had undergone one or multiple lines of treatment. Few MRIs with artifacts or MRIs in which the bone marrow could not delineated with high confidence were excluded, as we considered these datasets would not necessarily improve the algorithm. The flowchart is shown in Supplementary Figure 1.
Supplementary Figure 1. Flow chart for the development set.
Data set I: To test the algorithm, we chose consecutive patients with suspected or confirmed monoclonal plasma cell disorder undergoing wb-MRI at scanner 1 of center 1. We aimed to reach a number of at least 100 cases for Center 1, so the timespan from which data was included was set from 1.1.2020 to 31.12.2020 when searching for the respective wb-MRI examinations using a query in our radiological information system (RIS). From the respective query, wb-MRIs which were performed for other indications than suspected or confirmed
2
monoclonal plasma cell disorder were excluded. In order to perform a completely
independent test-set, patients which were also represented in the development set were also excluded. Then, examinations which had shown up in the query but which actually had not been performed, examinations which did not include a composed wb-ADC-map, and examinations which could not be retrieved from the PACs system were excluded.
Examinations with severe imaging artifacts in the pelvic area were also excluded. The flow- chart is shown in Supplementary Figure 2.
Supplementary Figure 2. *Volume of segmentation below cutoff. Abbreviations: wb: whole-
body, RIS: radiological information system, MPCD: monoclonal plasma cell disorder, PACS:
Picture archiving and communications system.
Dataset II: As a second test set for the algorithm, we chose consecutive patients with suspected or confirmed monoclonal plasma cell disorder undergoing wb-MRI at center 2 with scanner 3. We aimed to reach a number of at least 100 cases for center 2, so the timespan from which data was included was set from 1.1.2020 to 31.5.2020 when searching for the respective wb-MRI examinations using a query in our radiological information system 4
(RIS). From the respective query, wb-MRIs which were performed for other indications than suspected or confirmed monoclonal plasma cell disorder were excluded. In order to perform a completely independent test-set, patients which were also represented in the development set were also excluded. Then, examinations which had shown up in the query but which actually had not been performed, examinations which did not include a composed wb-ADC- map, and examinations which could not retrieved from the PACs system were excluded.
Examinations with severe imaging artifacts in the pelvic aera were also excluded. The flowchart is shown in Supplementary Figure 3.
Supplementary Figure 3. *Volume of segmentation below cutoff. Abbreviations: wb: whole-
body, RIS: radiological information system, MPCD: monoclonal plasma cell disorder, PACS:
Picture archiving and communications system.
Dataset III: As a third test set for the algorithm, we chose data from center 3 which was available at center 1 due to central image assessment for the GMMG-HD7 multicenter trial3 (EudraCT: 2017–004768-37). We used all datasets which were performed at scanner 4 or scanner 5. Wb-MRIs with missing composed wb-ADC-map and series with severe artifacts were excluded. Finally, 3 examinations were not successfully processed with the nnU-Net for segmentation. The flow-chart is shown in Supplementary Figure 4.
6
Supplementary Figure 4. *Volume of segmentation below cutoff. Abbreviations: wb: whole- body
Dataset IV: The intend of the third experiment was to investigate whether automatically extracted ADC values are correlated with established disease parameters: primarily, correlation with the plasma cell infiltration from bone marrow biopsy was investigated.
Therefore, we chose the baseline wb-MRIs from the GMMG-HD-7 trial3 (EudraCT: 2017–
004768-37) to use a cohort which had undergone both wb-MRI and bone marrow biopsy. As deviations in scanner or imaging protocol can lead to markedly different ADC-values,1 only examinations which had been performed at center 1 with scanner 1 with protocol 1
(Supplementary Figure 5 A) or at center 2 with scanner 3 with protocol 4 (Supplementary Figure 5 B) were included, as for these two scanners there was no bias detected in an earlier study.1 To receive a completely independent dataset from the development set, patients which were also represented in the development set were excluded. Wb-MRIs with other DWI protocols or without wb-ADC-map were excluded. The flow-chart is shown in
Supplementary Figure 5.
Supplementary Figure 5. Abbreviations: wb: whole-body
8
References:
1. Wennmann M, Thierjung H, Bauer F, et al. Repeatability and Reproducibility of ADC Measurements and MRI Signal Intensity Measurements of Bone Marrow in Monoclonal Plasma Cell Disorders. Invest. Radiol. 2022;57(4):272–281. Available at:
https://journals.lww.com/investigativeradiology/Fulltext/2022/04000/Repeatability_and_Re producibility_of_ADC.8.aspx.
2. Wennmann M, Bauer F, Klein A, et al. In Vivo Repeatability and Multi-Scanner Reproducibility of MRI Radiomics Features in Patients with Monoclonal Plasma Cell Disorders: A Prospective Bi-Institutional Study. Invest. Radiol. 2023;58(4):in press.
3. Goldschmidt H, Mai EK, Nievergall E, et al. Addition of Isatuximab to Lenalidomide, Bortezomib and Dexamethasone As Induction Therapy for Newly-Diagnosed, Transplant- Eligible Multiple Myeloma Patients: The Phase III GMMG-HD7 Trial. Blood.
2021;138(Supplement 1):463. Available at: https://doi.org/10.1182/blood-2021-145097.