سامانه همایش های دانشگاه آزاد اسلامی واحد نجف آباد

صفحه اصلی / International Conference on Artificial Intelligence; City, Industry and Health

Comparative Analysis of U-Net and U-Net (Xception) for CT-Based Segmentation of Target Volume and Organs At-Risk in Left Breast Cancer

نویسندگان :

Hajar Ahmadi¹ Azimeh NV Dehkordi² Farhad Azimifar³ Seied Rabi Mahdavi⁴ Mahnaz Roayaei⁵

1- Department of Biomedical Engineering, Isf.C., Islamic Azad University, Isfahan, Iran, ahmadierfane@yahoo.com 2- Department of Computer Engineering, Na.C., Islamic Azad University, Najafabad, Iran, nourizadeh@iau.ir 3- Department of Biomedical Engineering, Isf.C., Islamic Azad University, Isfahan, Iran, f.azimifar@khisf.ac.ir 4- 1. Department of Medical Physics and Radio-Oncology, Faculty of Medicine, Iran University of Medical Science, Tehran, Iran, srmahdavi@hotmail.com .2. d2 Radiation Biology Research Center, Iran University of Medical Science, Tehran, Iran, srmahdavi@hotmail.com 5- Department of Radiation Oncology, Omid Hospital, Esfahan University of Medical Sciences, Isfahan, Iran, roayaeimahnaz@gmail.com

کلمات کلیدی :

Breast Cancer،Radiotherapy treatment Planning،Segmentation،U-Net،Xception

چکیده :

Accurate segmentation of anatomical structures in breast cancer imaging is crucial for effective diagnosis and treatment planning. This study investigates the performance of U-Net and a novel U-Net variant that incorporates the Xception architecture as an encoder, aiming to enhance segmentation accuracy and efficiency. We conducted a comprehensive evaluation of both models using a diverse dataset, focusing on key metrics such as Dice similarity score, global accuracy, and prediction time. The U-Net model achieved a mean Dice score of 0.7434 with a global accuracy of 0.95109, demonstrating strong performance in segmenting structures such as the "Lung Left" and "Skin." However, the approach utilizing Xception significantly improved overall performance, yielding a mean Dice score of 0.78299 and a remarkable global accuracy of 0.98335, particularly excelling in clearly defined regions. both models facing challenges with the "GTV" class. This research underscores the potential of integrating advanced deep learning techniques in medical imaging, offering valuable insights for developing efficient tools to support oncological diagnostics and treatment strategies.