Related Experiment Video
Updated: Oct 3, 2025
10:44
Three-Dimensional Phase Resolved Functional Lung Magnetic Resonance Imaging
Published on: June 21, 2024
664
Anatomical Partition-Based Deep Learning: An Automatic Nasopharyngeal MRI Recognition Scheme.
Song Li1, Hong-Li Hua1, Fen Li2
1Department of Otolaryngology-Head and Neck Surgery, Renmin Hospital of Wuhan University, Wuhan, China.
Journal of Magnetic Resonance Imaging : JMRI
|February 14, 2022
Summary
This study developed a deep learning model using anatomical partitions for nasopharyngeal MRI disease recognition. The new method improved diagnostic accuracy, especially with limited data, offering a promising tool for clinical applications.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Otorhinolaryngology
Background:
- Deep learning (DL) model training for nasopharyngeal MRI disease recognition is challenging.
- Optimizing DL model performance requires innovative approaches.
Purpose of the Study:
- To develop an anatomical partition-based DL model integrating clinical anatomical knowledge for nasopharyngeal MRI disease recognition.
- To enhance the accuracy and efficiency of automated disease detection in nasopharyngeal imaging.
Main Methods:
- A retrospective study included 2485 patients with nasopharyngeal diseases and 600 controls.
- A segmentation model generated reduced-resolution images (seg112, seg224) from full-resolution nasopharyngeal MRI scans.
- Four pretrained DL models were trained on full and segmented datasets (100%, 50%, 25% data subsets).
Main Results:
- When using the full dataset (100%), models trained with seg112 and seg224 images showed similar performance (aAUC ~0.949) to those trained with full images (aAUC ~0.935).
- With a reduced dataset (25%), models trained with seg112 (aAUC 0.823) and seg224 (aAUC 0.765) images significantly outperformed models trained with full images (aAUC 0.640).
Conclusions:
- The proposed anatomical partition-based DL method shows potential for improving nasopharyngeal MRI disease recognition.
- This approach enhances DL model performance, particularly in scenarios with limited training data.
