Differentiation of eosinophilic and non-eosinophilic chronic rhinosinusitis on preoperative computed tomography using deep learning

Objective: This study aimed to develop deep learning (DL) models for differentiating between eosinophilic chronic rhinosinusitis (ECRS) and non-eosinophilic chronic rhinosinusitis (NECRS) on preoperative computed tomography (CT). Methods: A total of 878 chronic rhinosinusitis (CRS) patients undergoing nasal endoscopic surgery were included. Axial spiral CT images were pre-processed and used to build the dataset. Two semantic segmentation models based on U-net and Deeplabv3 were trained to segment sinus area in CT images. All patient images were segmented using the better-performing segmentation model and used for training and validation of the transferred efficientnet_b0, resnet50, inception_resnet_v2, and Xception neural networks. Additionally, we evaluated the performances of the models trained using each image and each patient as a unit. The precision of each model was assessed based on the receiver operating characteristic curve. Further, we analyzed the confusion matrix, accuracy, and interpretability of each model. Results: The Dice coefficients of U-net and Deeplabv3 were 0.953 and 0.961, respectively. The average area under the curve and mean accuracy values of the four networks were 0.848 and 0.762 for models trained using a single image as a unit, while the corresponding values for models trained using each patient as a unit were 0.853 and 0.893, respectively. The generated Grad-Cams showed good interpretability. Conclusion: Combining semantic segmentation with classification networks could effectively distinguish between patients with ECRS and NECRS based on preoperative sinus CT images. Furthermore, labeling each patient to build a dataset for classification may be more reliable than labeling each medical image.