Automatic bone marrow segmentation in whole-body magnetic resonance imaging: towards comprehensive, objective MRI-phenotypic bone marrow characterization in multiple myeloma

abstract

English

Background Whole-body magnetic resonance imaging (wb-MRI) is an important diagnostic tool for staging, risk assessment and response evaluation in myeloma. Wb-MRIs contain approximately 110 million voxels per sequence, and only a limited amount of this information can be processed and reported by radiologists to date. Deep learning has brought striking advances in biomedical image segmentation in recent years. The goal of this work was to establish an automatic whole-body bone marrow (BM) segmentation algorithm for T1-weighted MRI sequence, and to use these segmentations for comprehensive MRI-phenotypic characterization of the BM by subsequent radiomics analysis, bone by bone. Methods For 66 patients with smoldering multiple myeloma (SMM), BM was manually segmented on T1-w images. Thirty different BM compartments were individually labelled: right and left humerus, second to seventh vertebral bodies of the cervical spine (C2-C7), all vertebral bodies of the thoracic (T1-T12) and lumbar (L1-L5) spine, sacrum, right and left hip bone and right and left femur. Data was split by date by a 3:1 ratio into training-set and independent test-set. A nnU-Net, which is state of the art deep learning framework for medical image segmentation, was trained on the 52 training cases for segmentation of BM compartments, and postprocessing was used to distinguish sides of paired bones (humeri, hip bones, femora). Mean Dice scores report accuracy of automatic segmentations on the test-cases and interrater variability between two radiologists. This study was approved by the institutional review board. Results The mean Dice scores of the nnU-Net segmentation on the 14 test-cases for BM of right and left humerus, C2-C7, T1-T12, L1-L5, sacrum, right and left hip bone, right and left femur were 0.95, 0.94, 0.87, 0.86, 0.84, 0.80, 0.82, 0.84, 0.86, 0.87, 0.88, 0.91, 0.85, 0.85, 0.83, 0.83, 0.87, 0.91, 0.93, 0.93, 0.89, 0.85, 0.85, 0.81, 0.76, 0.88, 0.93, 0.93, 0.97 and 0.97, respectively. The mean Dice scores between segmentations from 2 radiologists on 2 cases for these BM compartments in the same order were 0.94, 0.95, 0.79, 0.86, 0.88, 0.81, 0.78, 0.79, 0.84, 0.81, 0.91, 0.85, 0.89, 0.90, 0.91, 0.91, 0.92, 0.90, 0.90, 0.89, 0.88, 0.92, 0.93, 0.88, 0.88, 0.81, 0.89, 0.89, 0.95, 0.94, respectively. On a descriptive level, we found differences in radiomics signatures between vertebrae with physiological bone marrow, vertebrae with focal lesions and vertebrae with diffuse infiltration in exemplary cases. Conclusion We established automatic, bone by bone BM segmentation in SMM patients with accuracy only slightly worse compared to the interrater variability of radiologists, mostly due to lumbosacral transitional vertebra. In exemplary cases we found different radiomics-signatures between physiological BM and different pathologies, indicating that such BM segmentations can be used for in depth BM characterization from wb-MRI when combined with subsequent radiomics analysis.

bone marrow phenotyping; bone marrow segmentation; MRI; multiple myeloma

WENNMANN, M.; CHMELÍK, J.; BAUER, F.; KLEIN, A.; UHLENBROCK, C.; LOCHNER, J.; GRÖZINGER, M.; ROTKOPF, L.; SAUER, S.; BONEKAMP, D.; KLEESIEK, J.; WEBER, T.; HILLENGASS, J.; GOLDSCHMIDT, H.; SCHLEMMER, H.; FLOCA, R.; WEINHOLD, N.; MAIER-HEIN, K.; DELORME, S.

1. 10. 2021

CIG MEDIA GROUP, LP

3500 MAPLE AVENUE, STE 750, DALLAS, TX 75219-3931

2152-2650

CL LYMPH MYELOM LEUK

21

11

unknown

S45

S46

1

https://reader.elsevier.com/reader/sd/pii/S2152265021021467?token=A5DB8D5CD125D2E691E636A9176F47540224719A63135FF6A50E181E4123DE4E3FB62E6C0BD343D7226E7B7CBA54567B&originRegion=eu-west-1&originCreation=20220524083008