Image Registartion And Object detection using Deep Learning In Radiotherapy

Abstract

Image registration and object detection are pivotal processes in the domain of radiotherapy, significantly enhancing the precision and effectiveness of treatment planning and delivery. This study leverages the capabilities of deep learning to address the challenges inherent in these tasks, particularly focusing on the complexities of anatomical variations and deformations. We propose and evaluate two deep learning models: VoxelMorph and a Sequential Model. VoxelMorph is tested on two datasets: the CTCBCT and Retinal FIRE datasets, achieving Cross Correlation (CC) values of 0.009 and 0.14, and Dice Similarity Coefficients (DiceSoccer) of 0.1 and 0.2, respectively. These results indicate that VoxelMorph demonstrates moderate performance in image registration, with better outcomes on the Retinal FIRE dataset compared to the CTCBCT dataset. The Sequential Model, designed for detecting and classifying tumor images in MRI datasets, demonstrates superior performance with a training loss of 0.191, training accuracy of 93%, validation accuracy of 98%, and test accuracy of 97%. These metrics underscore the model’s robustness and effectiveness in accurately identifying and classifying tumors. This study underscores the potential of deep learning, particularly convolutional neural networks (CNNs), in overcoming traditional challenges in image registration and object detection in radiotherapy.