Anatomical Partition-Based Deep Learning: An Automatic Nasopharyngeal MRI Recognition Scheme.

ABSTRACT

BACKGROUND: Training deep learning (DL) models to automatically recognize diseases in nasopharyngeal MRI is a challenging task, and optimizing the performance of DL models is difficult. PURPOSE: To develop a method of training anatomical partition-based DL model which integrates knowledge of clinical anatomical regions in otorhinolaryngology to automatically recognize diseases in nasopharyngeal MRI. STUDY TYPE Single-center retrospective study. POPULATION A total of 2485 patients with nasopharyngeal diseases (age range 14-82 years, female, 779[31.3%]) and 600 people with normal nasopharynx (age range 18-78 years, female, 281[46.8%]) were included. SEQUENCE 3.0 T; T2WI fast spin-echo sequence. ASSESSMENT Full images (512 × 512) of 3085 patients constituted 100% of the dataset, 50% and 25% of which were randomly retained as two new datasets. Two new series of images (seg112 image [112 × 112] and seg224 image [224 × 224]) were automatically generated by a segmentation model. Four pretrained neural networks for nasopharyngeal diseases classification were trained under the nine datasets (full image, seg112 image, and seg224 image, each with 100% dataset, 50% dataset, and 25% dataset). STATISTICAL TESTS The receiver operating characteristic curve was used to evaluate the performance of the models. Analysis of variance was used to compare the performance of the models built with different datasets. Statistical significance was set at P < 0.05. RESULTS: When the 100% dataset was used for training, the performances of the models trained with the seg112 images (average area under the curve [aAUC] 0.949 ± 0.052), seg224 images (aAUC 0.948 ± 0.053), and full images (aAUC 0.935 ± 0.053) were similar (P = 0.611). When the 25% dataset was used for training, the mean aAUC of the models that were trained with seg112 images (0.823 ± 0.116) and seg224 images (0.765 ± 0.155) was significantly higher than the models that were trained with full images (0.640 ± 0.154). DATA CONCLUSION: The proposed method can potentially improve the performance of the DL model for automatic recognition of diseases in nasopharyngeal MRI. LEVEL OF EVIDENCE 4 TECHNICAL EFFICACY STAGE 1.