Prediction of Chronic Rhinosinusitis With or Without Polyps Using Computed Tomography and Machine Learning

Introduction: Chronic rhinosinusitis (CRS) is clinically divided into two main phenotypes: CRS without nasal polyps (CRSsNP) and CRS with nasal polyps (CRSwNP). CRSwNP represents approximately 25-30% of patients and generally requires nasal endoscopy in for diagnosis. Thus far, there are no CT algorithms distinguishing these two differently managed phenotypes. Here, we present a pilot machine learning algorithm aimed at distinguishing CRSsNP from CRSwNP using a CT scan alone.

Methods: We utilized a random sample of a prospective cohort of CRS patients undergoing functional endoscopic sinus surgery (FESS). Patient CT scans were segmented by a multi-label fusion algorithm. in total, 7 CT scans were manually segmented as references into regions including all sinuses (bilateral maxillary, ethmoid, sphenoid, and frontal) and the nasal cavity. The scan was then nonlinearly registered with each of the 7 reference CTs and warped to generate labels. As features for a predictive model, 20-percentile value from each sinus, median values of central area of each sinus, and median values of left and right nasal cavity were extracted (n = 19). A random forest classifier was applied to predict the presence of NP, and 10-fold cross-validation were performed with accuracy as an outcome. As comparison, two otolaryngology residents reviewed the same CT images. Percent accuracy and interrater reliability were calculated.

Results: Of our 46-patient cohort, 31 (67%) of patients had NP and 15 (33%) did not. In the NP group, 61% were male (vs 33% in the group without NP, p = 0.14) and 87% were white (vs 73%, p = 0.44). The median [IQR] age of patients with NP was 47 [27, 57] (vs 30 [18, 53], p = 0.43) and median [IQR] SNOT-22 score at time of surgery was 47 [27, 57] (vs 46 [33.5, 62.5], p = 0.85). Using a random forest classifier to predict the presence of NP, a 10-fold cross-validation resulted in accuracy of 79%. Meanwhile, we found the human accuracy of distinguishing CRSwNP and CRSsNP using the same CT images to be 75% (κ=0.76).

Conclusions: This machine learning model demonstrates our ability to distinguish between CRSwNP and CRSsNP without nasal endoscopy.