Detecting abnormal thyroid cartilages on CT using deep learning.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Thyroid cartilage abnormalities can impact patient health and require accurate detection.
• Computed tomography (CT) is a key imaging modality for evaluating anatomical structures.
• Deep learning offers advanced computational approaches for medical image analysis.

Purpose of the Study:

• To assess the performance of a deep learning algorithm in identifying thyroid cartilage abnormalities.
• To evaluate the algorithm's ability to classify thyroid cartilage as normal or abnormal using CT data.

Main Methods:

• Utilized a dataset of 515 thyroid CT examinations, with 326 annotated for cartilage abnormalities.
• Employed transfer learning with the VGG16 neural network architecture for binary classification.
• Preprocessed CT images by cropping and data augmentation to enhance model training.

Main Results:

• The best-performing deep learning model achieved an area under the receiver operating characteristic curve (AUC) of 0.72 on test images.
• Sensitivity and specificity for abnormality detection were 83% and 64%, respectively, at the optimal threshold.
• An independent validation set of 189 images yielded an AUC of 0.70, confirming model generalizability.

Conclusions:

• Deep learning-based systems are feasible for evaluating thyroid cartilage abnormalities from CT scans.
• The developed model shows promise but does not yet match the diagnostic accuracy of human experts.
• Further refinement is needed to improve the algorithm's performance for clinical application.