1. Introduction
Chronic rhinosinusitis (CRS), which is characterized by inflammatory dysregulation of the nasal and parasinus mucosa for ≥12 consecutive weeks, is among the most common otolaryngological diseases, with a prevalence rate of 8% in the general population [1]. Currently, CRS is classified into two types based on the presence or absence of nasal polyps; further, it can be classified as eosinophilic sinusitis (ECRS) and non-eosinophilic sinusitis (NECRS) based on the eosinophilic infiltration level in the nasal mucosa or polyp. There has been recent progress in the elucidation of the pathogenesis of CRS.  There are several limitations in the traditional phenotypic classification of CRS into CRSwNP and CRSsNP. First, this classification does not account for pathophysiological differences; moreover, immune system cells are crucially involved in inflammatory mechanisms [2-4]. Many patients with extensive tissue eosinophil infiltration do not present polyp-like degeneration, with the incidence rate being approximately 27.5% [5]. Further, even patients with polypoid changes may predominantly present non-eosinophilic inflammation [6]. Classification of CRS into ECRS and NECRS is more likely to reflect the underlying inflammatory process. Additionally, the level of eosinophil infiltration in CRS lesions is strongly associated with postoperative outcomes [7, 8].
ECRS has a recurrence rate as high as 98.5%; additionally, it is the main cause of refractory sinusitis recurrence [9]. The gold standard for ECSR diagnosis is histopathological examination; however, it is invasive. Recently, the main treatments for ECSR and NECRS are contoured endoscopic nasal surgery and functional endoscopic nasal surgery, respectively [10, 11]. Given the high postoperative recurrence rate of ECRS and differences in the surgical methods for both CRS types, accurate preoperative prediction of the CRS type is necessary for predicting postoperative outcomes and administering personalized treatment. Eosinophil levels in the peripheral blood are associated with the eosinophil infiltration degree in the nasal sinus mucosa [12]. However, allergies, autoimmune diseases, drug reactions, parasitic infections, and corticosteroid treatment can alter circulating eosinophil levels. Additionally, increased eosinophil levels in peripheral blood do not necessarily reflect an increase in tissue eosinophils; moreover, the predictive utility of blood eosinophil levels for ECRS remains limited [13, 14]. There has been extensive research on the preoperative predictive utility of exhaled nitric oxide levels, serum total immunoglobulin E (IgE), specific IgE, and skin prick tests for ECRS and NECRS; however, they have low sensitivity and specificity [15, 16]. There are several differences in sinus computed tomography (CT) findings between patients with ECRS and NECRS. Early-stage ECRS often presents as ethmoid sinus lesions and mild maxillary sinus lesions, while NECRS often presents with maxillary sinus lesions. However, these characteristics do not effectively distinguish ECRS from NECRS [8, 17].
Considering the strong feature extraction and screening ability of artificial intelligence, applying artificial intelligence technology to sinus CT-assisted ECRS diagnosis may allow accurate preoperative prediction of ECRS. CRS lesions are limited to the nasal cavity and paranasal sinus areas in each disease type; moreover, the sinus area comprises a small part of the sinus CT images, with the surrounding tissue structure being complex. Therefore, noise information in the whole image increases the required training data. Manually delimiting the regions of interest (ROIs) is inconsistent with the original intention of artificial intelligence; further, its huge sketching workload greatly reduces its clinical application value. Therefore, we aimed to develop a segmentation model that allowed automatic segmentation of the sinus region. Moreover, since different networks have certain data ”preferences,” we used four common classification networks to train the segmented images.