SARA: Automated Anatomical Reference Assignment of Lower Limbs on MRI scans via ANN-based Bone Segmentation

Objective: The automated analysis of CT/MRI scans necessitates the preliminary processing of the collected data. A crucial step in this process is the unsupervised selection of slides within the region of interest. The objective of this study is to develop a data processing pipeline that employs a deep-learning (DL) approach to address this challenge. Methods: The proposed solution was implemented for the automated localisation of lower-limb bones on T1-weighted MRI scans. The model was trained using 1,243 MRI scans, with a subset of 29 manually labelled bone segmentations of six key lower-limb bone classes. The axial slices were divided into three sets: training (2,283), validation (300), and holdout (378). A two-part segmentation pipeline was developed using a combination of U-Net and ResNet architectures, with a custom cost function to handle variable label presence across slices. Results: Segmentation quality for the tibia and femur achieved Dice scores of 86.04% and 86.97% respectively, for the holdout subset. The ResNet classifier correctly identified the defined regions on the volume, achieving an area under the curve (AUC) of over 97% for the holdout subset for the majority of leg fragments, with the exception of the knee label. Conclusions: A novel DL-based pipeline for anatomical spatial localisation of MRI scans has been developed. The proposed methodology estimates the precise bone-related slice location within the scanned volume, thereby providing an added layer of anatomical context. This solution exhibits considerable potential for integration into clinical workflows, particularly in the context of studies pertaining to musculoskeletal diseases.