Multi-Class Segmentation of Organ at Risk from Thoracic Ct Images: A Deep Learning Approach

Abstract

Medical Imaging Segmentation is an essential technique for modern medical applications. It is the foundation of many aspects of clinical diagnosis, oncology and computer-integrated surgical intervention. Although significant successes have been achieved in the segmentation of medical images, deep learning approaches. Manual delineation of OARs is vastly dominant but it is prone to errors given the complex irregularities in shape, low texture diversity between tissues and adjacent blood area, patient-wide location of organisms, and weak soft tissue contrast across adjacent organs in CT images. Several models have been implemented in multi-class image segmentation up to this stage, but none of them address the issue of imbalanced classes, in which certain organs have relatively small pixels in comparison to others. To segment OARs in thoracic C T images, we proposed the model based on encoder decoder approach using transfer learning with efficientnetB7 deep learning model. We have built a fully connected Convolutional Neural Networks having 5 layers of encoding and 5 layers of decoding with efficientnetB7 specifically to tackle imbalance class pixels in an accurate way for the segmentation of OARs. Contrast enhanced CT is being used extensively for thoracic disorders and is of considerable significance in the medical field. The results showed that our proposed framework can be segmented organs accurately.