|
Enhancing medical image segmentation accuracy using a hybrid U-Net and ResNet18 architecture for improved treatment planning
|
Hajar Ahmadi1    , Azimeh NV Dehkordi2 , Farhad Azimifar3 , Seied Rabi Mehdi Mahdavi4 , Mahnaz Roayaei5 |
1- Department¬ of Biomedical Engineering, Isf.C., Islamic Azad University, Isfahan, Iran 2 Department of Physics, Na.C., Islamic Azad University, Najafabad, Iran
2- Assistant Professor, Department of Biomedical Engineering, Isf.C., Islamic Azad University, Isfahan, Iran, , nourizadeh@iau.ac.ir
3- Assistant Professor, Department of Biomedical Engineering, Isf.C., Islamic Azad University, Isfahan, Iran,
4- Department of Medical Physics and Radio-Oncology, Faculty of Medicine, Iran University of Medical Science, Tehran, Iran 5 Radiation Biology Research Center, Iran University of Medical Science, Tehran, Iran
5- Department of Radiation Oncology, Omid Hospital, Esfahan University of Medical Sciences, Isfahan, Iran |
|
|
Abstract: (3 Views) |
Background: Improving segmentation accuracy in CT images of patients with breast cancer plays a crucial role in designing effective treatment plans. This study aimed to evaluate the performance of two deep neural network models, including a U-Net architecture and a combined U-Net with ResNet-18, in this domain.
Materials and methods: Data from 120 patients, comprising 6,890 CT slices, were annotated by expert oncologists and used as the reference standard. These annotations included the heart, left lung, planning target volume (PTV), gross tumor volume (GTV), background, and skin. After data preprocessing, hyperparameters were optimized using a random search method. The models were trained and evaluated based on various performance metrics. Results indicated that the combined model outperformed the U-Net alone in identifying tumor volumes and sensitive structures.
Results: The Dice score reached to 60.35% for GTV and 76.94% for PTV, whereas the U-Net model achieved 39.22% and 39.47%, respectively. These findings demonstrated that the integrated architecture provided significant improvements in classification and detection accuracy, affirmed that combining these architectures enhanced segmentation performance.
Conclusion: Utilizing this approach can significantly improve treatment planning processes and reduce targeting errors. The results clearly show that the combined U-Net and ResNet-18 architecture performs better than the U-Net alone. However, further validation with larger datasets is necessary to establish its clinical applicability.
|
|
| Keywords: Breast cancer, Radiotherapy, Inverse treatment planning, Deep neural networks, U-Net, Resnet18. |
|
|
Full-Text [PDF 853 kb]
(1 Downloads)
|
Semi-pilot: Controlled/Randomized clinical trial |
Subject:
Medical Physics
Received: 2025/09/22 | Accepted: 2026/01/2 | Published: 2026/09/1
|
|
|
|
|
|
|
| Send email to the article author |
|
|
|
| Add your comments about this article |
|
|
|
|
Ahmadi H, NV Dehkordi A, Azimifar F, Mahdavi S R M, Roayaei M. Enhancing medical image segmentation accuracy using a hybrid U-Net and ResNet18 architecture for improved treatment planning. MEDICAL SCIENCES 2026; 36 (3) :289-301 URL: http://tmuj.iautmu.ac.ir/article-1-2433-en.html
|
