Automatic multi-structure pediatric knee bone segmentation using optimal multi-level Otsu thresholding to tackle intensity homogeneity in bone structures

Abstract

Recent studies in medical image segmentation involve new automatic approaches where active learning models are useful with less training samples. Presence of homogenous and heterogenous intensities for a single anatomical structure in pediatric musculoskeletal MR images affects the accuracy in terms of segmentation and classification of labels. This study addresses the homogeneity in intensity issues and introduces a new pre-training pipeline framework of Multi-level Otsu thresholding image as separate channel for 3D UNet model training. The proposed framework achieved higher performance of up to 85% when compared with the Baseline 3D UNet model and the Histogram threshold with 3D UNet. All algorithms are run through MONAI core framework.Clinical Relevance- This study will be of major interest to practicing pediatric clinicians and surgeons for its ability to provide accurate morphological assessment of underlying musculoskeletal structure. For researchers, it provides a new approach in dealing with heterogeneity in intensity problem which is common in pediatric MR imaging.