Artificial intelligence driven diagnostic model for detecting paranasal sinus opacification in computed tomography...

Visual analysis of paranasal sinuses (PNS) on computed tomography (CT) images requires interpretation and reporting of sinus involvement and other anatomical factors. This is time consuming, labour intensive and subjective. Artificial intelligence (AI)-based machine learning (ML) tools are under development for analysis of radiological images. The scope of this study was to develop and evaluate a coding-free ML model for automated identification of PNS on CT images.

A total of 19,119 anonymous coronal images retrieved from 90 CT studies were included. All images were annotated with locations, names and opacification status of the sinuses. The images were divided into training, validation and testing datasets. The ML model was trained for 2000 iterations using YOLOv2 algorithm, and its accuracy was evaluated using F1 score and Intersection over Union (IoU) metrics.

An ML model was developed using "Create ML" application on an Apple MacBook computer. A mean F1 score of 0.89 and a mean IoU50 of 79% was achieved during evaluation of the model on the testing dataset. The highest accuracy was seen in the detection of normal sphenoid sinus, and the lowest in the detection of opacified frontal sinus.

The study demonstrates the utility of AI and ML in automating the interpretation of PNS CT images. From the results of our study, it can be concluded that a coding-free ML model can be developed and deployed for automated identification of PNS on CT images with accuracy similar to custom-coded ML models.