Automated FDG uptake/PET-CT fused scan diagnosis of various lymph node tumors using object detection AI techniques.
Authors
Affiliations (5)
Affiliations (5)
- Physics, Faculty of Science, Mansoura University, 35516, Mansoura, Dakahlia, Egypt.
- Information Technology, Faculty of Computers and Information, Arish University, 45511, Al-Arish, North Sinai, Egypt.
- Computers and Systems, Electronics Research Institute, Cairo, 12622, Egypt.
- Physics, Faculty of Sciences, Bisha University, 61922, Bisha, Saudi Arabia.
- Information Technology, Faculty of Computers and Information, Mansoura University, 35516, Mansoura, Dakahlia, Egypt. [email protected].
Abstract
Lymph nodes (LN) constitute a vital component of the lymphatic system, serving a pivotal role in immune functioning and maintaining fluid balance in the body. Moreover, they serve as markers for tailoring treatments. Inaccurate assessment of LN status may lead to either inadequate treatment or an overly aggressive treatment approach, thereby heightening the risk of recurrence and postoperative complications. Many imaging techniques are used to assess and characterize LNs, but they are limited by low sensitivity for detecting small metastases. Therefore, artificial intelligence (AI) object detection techniques are utilized to localize the relevant objects in images and classify them into relevant classes. This paper proposes a method for detecting 13 LN classes across different body organs using real-world PET-CT datasets. We provide two modules: the first establishes a new LN dataset by fusing CT and PET images for each patient, then denoising and annotating the classes. The dataset was first divided into 80% training, 10% validation, and 10% testing, with data augmentation applied only to the training set to avoid data leakage. Subsequently, 5-fold cross-validation was conducted on the training and validation data to ensure a more reliable evaluation. The final results are reported based on the cross-validation protocol, while the hold-out test set is used for independent assessment. The second module is object detection based on a modified YOLOv8. We selected the kernel, optimized the feature-extraction backbone layers, and tuned other hyperparameters. We compared the performance of eight popular one-stage architectures: YOLOv7, YOLOv8, YOLOv9, YOLOv10, YOLOv11, YOLOv12, YOLONas, and the modified YOLOv8. Work performance has been measured using precision, recall, mean average precision (mAP50), and Dice similarity coefficient (DSC). The findings demonstrate the superiority of the proposed method, with improvements of 78%, 75%, 81%, and 76%, respectively.