Diagnosis of skull-base invasion by nasopharyngeal tumors on CT with a deep-learning approach

TL;DR

A deep-learning model was developed to diagnose skull-base invasion by nasopharyngeal tumors in CT scans, outperforming radiologists in accuracy.

Contribution

A novel CNN model using transfer learning with ResNet-50 for diagnosing skull-base invasion in CT images of nasopharyngeal tumors.

Findings

The CNN model achieved a diagnostic accuracy of 0.973, significantly higher than experienced and junior radiologists.

Grad-CAMs showed that the model focused on bone marrow for invasion-negative cases and osteosclerosis/masses for invasion-positive cases.

The model's area under the curve (0.953) was significantly better than both radiologists (0.832 and 0.617).

Abstract

To develop a convolutional neural network (CNN) model to diagnose skull-base invasion by nasopharyngeal malignancies in CT images and evaluate the model’s diagnostic performance.

We divided 100 malignant nasopharyngeal tumor lesions into a training (n = 70) and a test (n = 30) dataset. Two head/neck radiologists reviewed CT and MRI images and determined the positive/negative skull-base invasion status of each case (training dataset: 29 invasion-positive and 41 invasion-negative; test dataset: 13 invasion-positive and 17 invasion-negative). Preprocessing involved extracting continuous slices of the nasopharynx and clivus. The preprocessed training dataset was used for transfer learning with Residual Neural Networks 50 to create a diagnostic CNN model, which was then tested on the preprocessed test dataset to determine the invasion status and model performance. Original CT images from the test dataset were reviewed by a radiologist with extensive head/neck imaging experience (senior reader: SR) and another less-experienced radiologist (junior reader: JR). Gradient-weighted class activation maps (Grad-CAMs) were created to visualize the explainability of the invasion status classification.

The CNN model’s diagnostic accuracy was 0.973, significantly higher than those of the two radiologists (SR: 0.838; JR: 0.595). Receiver operating characteristic curve analysis gave an area under the curve of 0.953 for the CNN model (versus 0.832 and 0.617 for SR and JR; both p < 0.05). The Grad-CAMs suggested that the invasion-negative cases were present predominantly in bone marrow, while the invasion-positive cases exhibited osteosclerosis and nasopharyngeal masses.

This CNN technique would be useful for CT-based diagnosis of skull-base invasion by nasopharyngeal malignancies.

Figures