Building a Large Comprehensive Medical Image Set of Sinus Diseases

Abstract

Sinuses disorders are among the most common disorders that affect people’s lives worldwide. Diagnosing such disorders requires highly skilled specialists to carefully inspect Computed Tomographic (CT) scans of the patient. The diagnosis process is time-consuming and very costly. To build a machine learning based computer system for the diagnosis process, an annotated set of CT scans representing different sinus disorders is needed to train and test such a system. In this work, we build an image set by collecting CT scans of 100 patients with an average of 94 slices per patient. In each scan, ten different sinuses and sinus parts are captured. These sinuses and sinus parts are distinguished as Frontal (right side), Frontal (left side), Maxillary (right side), Maxillary (left side), Anterior Ethmoid (right side), Anterior Ethmoid (left side), Posterior Ethmoid (right side), Posterior Ethmoid (left side), Sphenoid (right side), and Sphenoid (left side). The scans are segmented and annotated by specialists, where each segment is labeled with the sinus (or sinus part) it depicts (one out of the ten classes mentioned above) along with one of the following six classes representing the status of this part: Normal, Cyst, Osteoma, Chronic Rhinosinusitis (CRS), Antrochoanal polyp (ACP), and Missing sinus. The dataset is acquired from the King Abdullah University Hospital (KAUH) in Jordan and it consists of 48,324 different annotated samples making it the largest and most comprehensive dataset for sinus diseases to the best of our knowledge.