A Deep Learning aided bone marrow segmentation of quantitative fat MRI for myelofibrosis patients

Humera Tariq^*Lubomir HadjiiskiDariya MalyarenkoMoshe TalpazKristen PettitGary D LukerBrian D RossThomas L Chenevert^*

• University of Michigan, Ann Arbor, United States

To automate bone marrow segmentation within pelvic bones in quantitative fat MRI of myelofibrosis (MF) patients using deep-learning (DL) U-Net models.Methods: Automated segmentation of bone marrow (BM) was evaluated for four U-Net models: 2D U-Net, 2D attention U-Net (2D A-U-Net), 3D U-Net and 3D attention U-Net (3D A-U-Net). An experienced annotator performed the delineation on in-phase (IP) pelvic MRI slices to mark the boundaries of BM regions within two pelvic bones: proximal femur and posterior ilium. The dataset comprising volumetric images of 58 MF patients was split into 32 training, 6 validation and 20 test sub-sets. Model performance was assessed using conventional metrics: average Jaccard Index (AJI), average Volume Error (AVE), average Hausdorff Distance (AHD), and average Volume Intersection Ratio (VIR). Iterative model optimization was performed based on maximizing validation sub-set AJI. Wilcoxon's rank sum test with Bonferroni corrected significance threshold of p<0.003 was used to compare DL segmentation models for test sub-set.Results: 2D segmentation models performed best for iliac BM with achieved scores of 95-96% for the VIR and 87-88% for AJI agreement with expert annotations on the test set. Similar performance was observed for femoral BM segmentation with slightly better VIR but worth AJI agreement for U-Net (94% and 86%) versus A-U-Net (92% and 87%). 2D models also exhibited lower AVE variability (8-9%) and ilium AHD (16mm). The 3D segmentation models have shown marginally higher errors (AHD of 19-20mm for ilium and 10-12% AVE-SD for both bones) and generally lower agreement scores (VIR of 91-93% for ilium and 89-91%for femur with 85-86% AJI).