Article Dans Une Revue Asian Journal of Medicine and Health Année : 2025

Automatic Segmentation of Organ at Risk in Head and Neck Cancer CT Images Using Medical Open Network for Artificial Intelligence (MONAI) with Deep Learning Techniques

Résumé

Aims: This study aims to evaluate the deep learning-based Medical Open Network for Artificial Intelligence (MONAI) framework in performing automatic segmentation of Organ at risk (OAR) structures, especially in cases of head and neck cancer. Study Design: This research an experimental design with a quantitative approach to predict the segmentation of annotated data in head and neck cancer using the U-Net model. The study involved training and testing using data from The Cancer Imaging Archive (TCIA) and a publicly available dataset, evaluating its performance based on balanced accuracy, precision, recall, and Dice Similarity Coefficient (DSC). The result was analyzed to determine the model accuracy in replicating actual anotated segmentation, providing a data driven assessment of its predictive capability. Place and Duration of Study: This research was conducted in Department of Physics Study Program, Udayana University, between January 2024 and June 2025. Methodology: The dataset with 179 subjects consist of two dataset sources, namely the HEAD-NECK-RADIOMICS-HN1 (H&N1) Collection and the Han-Seg collection with CT image sizes of 512 x 512 x 116 - 512 x 512 x 323. The model trains and tests using that annotated dataset, focusing on nine organ labels: brainstem, right submandibular, left submandibular, mandible, optic chiasm, optic nerve, right parotid, left parotid, and spinal cord. The U-Net model was implemented using the MONAI framework with preprocessing stages of converting the dataset to a NIfTI file, creating a dataset.json file, label remapping and applying MONAI transforms, namely deterministic and augmentation. The U-Net model evaluation is performed using balanced accuracy, precision, recall, and DSC. Results: The model showed good segmentation performance with an average balanced accuracy value of 0.932 ± 0.043, precision of 0.690 ± 0.189, and recall of 0.863 ± 0.086. The precision values are noticeably lower compared to recall and balanced accuracy. The lowest precision among all OARs was observed for the optic chiasm at 0.353, while the highest was for the brainstem at 0.904. The mean value of Dice Similarity Coefficient (DSC) is 0.753 ± 0.141. The highest Dice score obtain in the mandible organ at 0.902, while the lowest value obtain in the optic chiasm at 0.500. The visualization of the results supports the finding that segmentation is more accurate on large and clear organs. Conclusion: The MONAI framework is able to perform automatic OAR segmentation with promising results, especially on large organs. However, performance on small organs is still low due to voxel limitations and label imbalance.

Fichier non déposé

Dates et versions

hal-05170629 , version 1 (19-07-2025)

Identifiants

  • HAL Id : hal-05170629 , version 1

Citer

I G.A Widyartha, Ni Nyoman Ratini, A.A Gunawan, Gusti Ngurah Sutapa, Ni Made Yuliara, et al.. Automatic Segmentation of Organ at Risk in Head and Neck Cancer CT Images Using Medical Open Network for Artificial Intelligence (MONAI) with Deep Learning Techniques. Asian Journal of Medicine and Health, 2025, 23 (7), pp.146-155. ⟨hal-05170629⟩
14 Consultations
0 Téléchargements

Partager

  • More