Combining Deep Learning and Radiomics for Automated, Objective, Comprehensive Bone Marrow Characterization From Whole-Body MRI

Abstract

Objectives: Disseminated bone marrow (BM) involvement is frequent in multiple myeloma (MM). Whole-body magnetic resonance imaging (wb-MRI) enables to evaluate the whole BM. Reading of such whole-body scans is time-consuming, and yet radiologists can transfer only a small fraction of the information of the imaging data set to the report. This limits the influence that imaging can have on clinical decision-making and in research toward precision oncology. The objective of this feasibility study was to implement a concept for automatic, comprehensive characterization of the BM from wb-MRI, by automatic BM segmentation and subsequent radiomics analysis of 30 different BM spaces (BMS). Materials and Methods: This retrospective multicentric pilot study used a total of 106 wb-MRI from 102 patients with (smoldering) MM from 8 centers. Fifty wb-MRI from center 1 were used for training of segmentation algorithms (nnU-Nets) and radiomics algorithms. Fifty-six wb-MRI from 8 centers, acquired with a variety of different MRI scanners and protocols, were used for independent testing. Manual segmentations of 2700 BMS from 90 wb-MRI were performed for training and testing of the segmentation algorithms. For each BMS, 296 radiomics features were calculated individually. Dice score was used to assess similarity between automatic segmentations and manual reference segmentations. Results: The "multilabel nnU-Net" segmentation algorithm, which performs segmentation of 30 BMS and labels them individually, reached mean dice scores of 0.88 ± 0.06/0.87 ± 0.06/0.83 ± 0.11 in independent test sets from center 1/center 2/center 3-8 (interrater variability between radiologists, 0.88 ± 0.01). The subset from the multicenter, multivendor test set (center 3-8) that was of high imaging quality was segmented with high precision (mean dice score, 0.87), comparable to the internal test data from center 1. The radiomic BM phenotype consisting of 8880 descriptive parameters per patient, which result from calculation of 296 radiomics features for each of the 30 BMS, was calculated for all patients.Exemplary cases demonstrated connections between typical BM patterns in MM and radiomic signatures of the respective BMS. In plausibility tests, predicted size and weight based on radiomics models of the radiomic BM phenotype significantly correlated with patients' actual size and weight (P = 0.002 and P = 0.003, respectively). Conclusions: This pilot study demonstrates the feasibility of automatic, objective, comprehensive BM characterization from wb-MRI in multicentric data sets. This concept allows the extraction of high-dimensional phenotypes to capture the complexity of disseminated BM disorders from imaging. Further studies need to assess the clinical potential of this method for automatic staging, therapy response assessment, or prediction of biopsy results.

In Vivo Repeatability and Multiscanner Reproducibility of MRI Radiomics Features in Patients With Monoclonal Plasma Cell Disorders

“…13 Few studies have applied radiomics to analyze MRIs from multiple myeloma patients so far. [39][40][41][42] Knowledge on robust RFs in multiple myeloma could be of use to reduce the RF set before modeling, especially given the limited cohort sizes of these studies, and help to improve the robustness of radiomics models for multicentric application.…”

Deep Learning for Automatic Bone Marrow Apparent Diffusion Coefficient Measurements From Whole-Body Magnetic Resonance Imaging in Patients With Multiple Myeloma

“…It produced segmentations with mean dice scores of 0.92 and 0.93 for the pelvic bones and 0.85 for the sacral bone, which were even higher than the dice scores of the interrater experiments that were performed to set a benchmark. The dice scores of the ADC-nnU-Net are similar to the dice scores for bone marrow segmentation from T1-weighted wb-MRI images from 1 prior study 28 and markedly higher than the dice scores for bone segmentation from another wb-MRI study that reported dice scores of 0.76 for the best achieved algorithm 37 . The quality for the segmentation of the sacral bone was lower than the quality for the segmentation of the pelvic bones, which is in line with the lower interrater reproducibility of the sacrum in this study, and is also in line with findings from automatic bone marrow segmentations on T1-weighted images 28 .…”

“…The dice scores of the ADC-nnU-Net are similar to the dice scores for bone marrow segmentation from T1-weighted wb-MRI images from 1 prior study 28 and markedly higher than the dice scores for bone segmentation from another wb-MRI study that reported dice scores of 0.76 for the best achieved algorithm 37 . The quality for the segmentation of the sacral bone was lower than the quality for the segmentation of the pelvic bones, which is in line with the lower interrater reproducibility of the sacrum in this study, and is also in line with findings from automatic bone marrow segmentations on T1-weighted images 28 . The quality for the segmentation was lower in data set 3 from center 3, which can be explained by the fact that the nnU-Net was trained on data from center 1 and 2, and therefore performed better on data from these scanners.…”

“…The default nearest neighbor interpolation between slices was changed to a linear interpolation because it yielded better results in our case. The right and left pelvis were relabeled by a postprocessing step, which is based on a mutual position between the sacral bone centroid and the centroid of the processed bone, as presented elsewhere 28 . A plausibility check on the automatically received segmentations was performed, in which pelvic bones with a volume below 100 cm 3 and sacral bones with a volume below 60 cm 3 were considered to be not successfully segmented, and these were excluded from further analysis.…”

Weakly Supervised Learning with Positive and Unlabeled Data for Automatic Brain Tumor Segmentation

“…While there are methods to control for the influence of rater variability from the segmentations [45,46], the usage of automatic segmentations is often considered better as the transfer to other clinical settings is more straightforward [47]. While there are some studies that make use of automatic segmentations [48,49], this is not always true. We assume that the comparatively small size of many studies [16,50,51] hinders the researcher in investing additional efforts.…”