HCIU: Hybrid clustered inception‐based UNET for the automatic segmentation of organs at risk in thoracic computed tomography images

Abstract

AbstractAutomatic segmentation is the most effective and fast method to plan radiotherapy treatment in thoracic computed tomography (CT), as manual segmentation takes time and produces variability in the results. This article proposes a hybrid HCIU method to segment the thoracic organs at risk (OAR) using prior localization techniques. The dataset of 40 Patients containing 7420 slices of thoracic CT has been used to perform the experiment on the Google Colab Pro+ framework. HCIU constitutes a hybrid deep learning model for features extraction, a K‐mean algorithm to form the clusters based on the extracted features, and the ensemble UNET InceptionV3 model trained on the data of each cluster for the segmentation of OAR. With HCIU, we achieved an average dice score of 92.05%, 97.39%, 95.75%, and 96.45% for the esophagus, heart, trachea, and aorta. The average sensitivity and specificity values achieved were 98.38% and 97.43% on the test dataset. Also, statistical tests have been applied to each organ's pixel values for analyzing the model's segmented results. A hybrid method has been proposed to improve the localization in thoracic OAR segmentation in CT. The proposed method achieved precise accuracy in segmenting the four organs compared with the other state‐of‐the‐art methods.