Automated segmentation of magnetic resonance bone marrow signal: a feasibility study

Abstract

Background Manual assessment of bone marrow signal is time-consuming and requires meticulous standardisation to secure adequate precision of findings. Objective We examined the feasibility of using deep learning for automated segmentation of bone marrow signal in children and adolescents. Materials and methods We selected knee images from 95 whole-body MRI examinations of healthy individuals and of children with chronic non-bacterial osteomyelitis, ages 6–18 years, in a longitudinal prospective multi-centre study cohort. Bone marrow signal on T2-weighted Dixon water-only images was divided into three color-coded intensity-levels: 1 = slightly increased; 2 = mildly increased; 3 = moderately to highly increased, up to fluid-like signal. We trained a convolutional neural network on 85 examinations to perform bone marrow segmentation. Four readers manually segmented a test set of 10 examinations and calculated ground truth using simultaneous truth and performance level estimation (STAPLE). We evaluated model and rater performance through Dice similarity coefficient and in consensus. Results Consensus score of model performance showed acceptable results for all but one examination. Model performance and reader agreement had highest scores for level-1 signal (median Dice 0.68) and lowest scores for level-3 signal (median Dice 0.40), particularly in examinations where this signal was sparse. Conclusion It is feasible to develop a deep-learning-based model for automated segmentation of bone marrow signal in children and adolescents. Our model performed poorest for the highest signal intensity in examinations where this signal was sparse. Further improvement requires training on larger and more balanced datasets and validation against ground truth, which should be established by radiologists from several institutions in consensus.

A novel deep learning method for large-scale analysis of bone marrow adiposity using UK Biobank Dixon MRI data

“…Several other recent studies have developed deep learning for automated BM segmentation from MRI data. For example, von Brandis et al assessed the feasibility of deep learning for segmenting BM from T2-weighted Dixon water-only images, focusing on the knee region [22] ; however, the best median dice score of their model was only 0.68, far below that obtained by our models ( Table 2 ). Better accuracy was achieved by Zhou et al, who established a deep learning model for segmenting lumbar vertebrae from Dixon MRI data [20] .…”

“…Firstly, our models were trained and tested using manual segmentations from only a single reader. In contrast, two previous BM segmentation models were trained and tested using manual segmentations produced by two independent human readers [20] , [22] ; this multi-reader approach can help to ensure consistency in the ground truths. However, single-reader ground truths have also been used to successfully develop other recent deep learning models for bone or BM segmentation [21] , [45] , and our ground truths were produced by a reader with extensive experience.…”

“…However, this study used only T1-weighted MR data and attempted to quantify BM adiposity based on raw MRI signal intensity, which has never been validated for this purpose [19] ; the clinical significance of calvarial BM adiposity also remains uncertain. Machine learning has also recently been used to segment the knee or vertebral BM from Dixon images in smaller cohorts outwith the UKBB [20] , [21] , [22] ; however, there are no peer-reviewed studies establishing machine learning for automated segmentation of the BM from other skeletal sites, and never using MR data from the UKBB. These were the goals of the present study.…”

Repeatability and test–retest reproducibility of mean apparent diffusion coefficient measurements of focal and diffuse disease in relapsed multiple myeloma at 3T whole body diffusion-weighted MRI (WB-DW-MRI)

“…However, this is the approach taken in previous myeloma studies 14 since unlike other primary tumours, myeloma lesions tend to be numerous so a single slice approach is less time consuming which may be more feasible for potential future clinical use. However, future studies may assess whole tumour segmentation tools using machine learning algorithm to measure disease burden and assess response which the current evidence supports its feasibility 26,27 This is currently under ongoing research in our institute (Machine Learning in Myeloma Response (MALIMAR) study. 28…”

A novel deep learning method for large-scale analysis of bone marrow adiposity using UK Biobank Dixon MRI data

“…Several groups have developed machine learning for automated segmentation of other anatomical regions from the UKBB MRI data (15)(16)(17). Machine learning has also recently been used to segment the knee or vertebral BM from Dixon images in smaller cohorts outwith the UKBB (18)(19)(20); however, machine learning has not yet been developed for automated segmentation of the BM from other skeletal sites, and never using MR data from the UKBB. These were the goals of the present study.…”