Automatic analysis of magnetic resonance imaging in multiple myeloma patients: deep-learning based pelvic bone marrow segmentation and radiomics analysis for prediction of plasma cell infiltration

abstract

English

Background Advances in deep learning have made automatic biomedical image segmentation feasible. Additionally, radiomics analysis now allows computer based, in-depth tissue analysis from medical images. The goal of this work was to establish a full-automatic framework combining automatic pelvic bone marrow (BM) segmentation and radiomics analysis of the pelvic BM to predict BM plasma cell infiltration (PCI) directly and automatically from whole-body magnetic resonance imaging (wb-MRI). Methods A total of 541 MRIs acquired at 5 different MRI scanners from 270 patients with all stages of monoclonal plasma cell disorders were included. One-hundred fifty-eight patients who had received MRI at the standard clinical 1.5T MRI scanner and had information on PCI from concomitant BM biopsy at the iliac crest available were split by date into a training set (n=116) for both, nnU-Net and radiomics model, and an independent test set for the framework (n=42). All MRIs without biopsy data were used for training of the nnU-Net only, which is state of the art deep learning framework for medical image segmentation. Manual BM segmentations of the right and left pelvic bone on coronal T1tse images on the training cohort were used for training of the nnU-Net. A random forest classifier was trained on the pelvic radiomics features to predict PCI. The framework was then tested on the independent test set. Dice scores report accuracy of segmentations. Mean absolute error (MAE) in [%PCI] reports the accuracy of the PCI prediction. For comparison, two radiologists rated the diffuse infiltration according to 3 levels of severity (none-to-mild vs. moderate vs. severe). The mean PCI within each severity level from the training set was determined and assigned as a prediction to patients with the same level diffuse infiltration in the test set (the radiologists’ PCI prediction). This study was approved by the institutional review board. Results The mean Dice scores of the nnU-Net segmentation for right / left pelvic BM on 8 cases of the test set (last 8 by acquisition date) were 0.94 and 0.94, respectively, the mean Dice scores between manual pelvic BM segmentation of two radiologists on these 8 cases were 0.87 and 0.88. The MAE of the prediction of PCI by the automated framework was 14.3 [%PCI]. The MAE of the radiologists PCI predictions were 16.1 [%PCI] (rater 1) and 16.7 [%PCI] (rater2). Conclusion We established automatic pelvic BM segmentation with radiologist level precision for all stages of monoclonal plasma cell disorders, which is a crucial step towards fully automated analysis of wb-MRI. Radiomics analysis of these segmentations can predict PCI with considerable accuracy. Further improvement of the PCI prediction model is necessary and is currently in progress, by adding additional MRI sequences into the model, enhancing the amount of training data by adding multi-institutional data, and learning about multi-scanner radiomics feature stability.

radiomics; deep learning; MRI; plasma cell infiltration; multiple myeloma

WENNMANN, M.; KLEIN, A.; BAUER, F.; CHMELÍK, J.; UHLENBROCK, C.; GRÖZINGER, M.; ROTKOPF, L.; SAUER, S.; THIERJUNG, H.; 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

S49

S49

1

https://reader.elsevier.com/reader/sd/pii/S2152265021021522?token=4354AD84998FC3D2A06FBF97906D1AA1A82AA6888682A33B7E3F3A6E2179F2F5E30D2CC5CB2CC85E4CADD598E96AC8F4&originRegion=eu-west-1&originCreation=20220524082854