Label-Free Machine Learning-Based Segmentation of Whole-Body Bone Marrow Imaging in Multiple Myeloma

Multiple myeloma is a plasma cell neoplasm with genetic complexity that originates in pre-malignant stages due to genomic alterations, leading to malignant plasma cell proliferation. Monitoring alterations in bone marrow during treatment for multiple myeloma may optimize patient outcomes. Multiple myeloma is a neoplasm that can distort bones, compromising the efficacy of conventional segmentation methodologies during the evaluation of whole body low dose computed tomography. A combination of pre-trained deep learning and traditional image analysis techniques for wholebody bone marrow segmentation is explored in this study. The pre-trained TotalSegmentator model was used to segment the entire skeleton in order to isolate key regions of interest and reduce pixel imbalances between the background and the bone pixels. A clustering technique was adapted to identify the bone marrow pixels and morphological operations were employed to refine the segmentation mask. The qualitative analysis performed by experienced radiologists shows promising results. The proposed segmentation pipeline allows accurate and fast annotation of the whole-body bone marrow imaging in multiple myeloma patients, achieving an IoU of $0.79 \pm 0.05$ on the available cohort with femur bone annotations.