CT Dataset Enhancement using Additional Feature Insertion for Automatic Femur Segmentation Model Based on Deep Learning

Abstract

This paper proposed to insert additional feature into input datasets (i.e., CT scans) for automatic femur segmentation model, U-Net, with respect to increase the accuracy of model performance. An additional feature is available reference information representing identity on each CT scans and has an effect on results of deep learning model training. In this experiment, choose the left-femur as the target organ, which is common organs-at-risk (OARs) for lower abdominal cancers. The automatic femur segmentation model training was separately executed through two different datasets, one cropped-dataset with additional feature and one original dimension dataset without additional feature. For additional feature, lying posture of patient when entered the CT scanner was selected. The performance results of both trained U-Net models were compered in order to observe the difference of effect. Evaluation results reported that the additional feature could increase an accuracy and precision including support prediction for the left-femur segmentation, with the Dice Similarity Coefficient (DSC) of 61.573% and Intersection Over Union (IoU) of 45.621%, respectively. Specifically, deep learning combining additional feature insertion on cropped-datasets was the novelty in this experiment to effectively segment the left femur.