Deep learning detects retropharyngeal edema on MRI in patients with acute neck infections

Abstract

Background: In acute neck infections, magnetic resonance imaging (MRI) shows retropharyngeal edema (RPE), which is a prognostic imaging biomarker for a severe course of illness. This study aimed to develop a deep learning-based algorithm for the automated detection of RPE.

Methods: We developed a deep neural network consisting of two parts using axial T2-weighted water-only Dixon MRI images from 479 patients with acute neck infections annotated by radiologists at both slice and patient levels. First, a convolutional neural network (CNN) classified individual slices; second, an algorithm classified patients based on a stack of slices. Model performance was compared with the radiologists' assessment as a reference standard. Accuracy, sensitivity, specificity, and area under receiver operating characteristic curve (AUROC) were calculated. The proposed CNN was compared with InceptionV3, and the patient-level classification algorithm was compared with traditional machine learning models.

Results: Of the 479 patients, 244 (51%) were positive and 235 (49%) negative for RPE. Our model achieved accuracy, sensitivity, specificity, and AUROC of 94.6%, 83.3%, 96.2%, and 94.1% at the slice level, and 87.4%, 86.5%, 88.2%, and 94.8% at the patient level, respectively. The proposed CNN was faster than InceptionV3 but equally accurate. Our patient classification algorithm outperformed traditional machine learning models.

Conclusion: A deep learning model, based on weakly annotated data and computationally manageable training, achieved high accuracy for automatically detecting RPE on MRI in patients with acute neck infections.

Relevance statement: Our automated method for detecting relevant MRI findings was efficiently trained and might be easily deployed in practice to study clinical applicability. This approach might improve early detection of patients at high risk for a severe course of acute neck infections.

Key points: Deep learning automatically detected retropharyngeal edema on MRI in acute neck infections. Areas under the receiver operating characteristic curve were 94.1% at the slice level and 94.8% at the patient level. The proposed convolutional neural network was lightweight and required only weakly annotated data.

Keywords: Artificial intelligence; Magnetic resonance imaging; Neural networks (computer); Respiratory tract infections; Retropharyngeal abscess.

Fig. 1

Examples of retropharyngeal edema (arrows) on axial T2-weighted Dixon water-only images of four different patients (a, b, c, d) with acute neck infections

Fig. 2

Study overview

Fig. 3

The architecture of our convolutional neural network with all the layers and their dimensions visible

Fig. 4

The structure of our proposed method to obtain the final classifications of a single patient: First, the convolutional neural network (CNN) is used to obtain numeric slice-wise predictions, then the maximum of the mean values of five consecutive slice-wise predictions is computed, and the patient-wise classification is obtained based on whether the maximum is below or above a certain threshold. If the patient is classified as positive, the slice-wise predictions are converted into binary labels with a second threshold; otherwise, they are all classified as negative. The two thresholds are chosen by finding the thresholds that give the maximal Youden’s index for the patient- and slice-wise predictions with this same method for the training set predictions. Our post-processing algorithm finds the appropriate thresholds and converts the predictions of all the test set patients into final binary labels

Fig. 5

The median AUROC curves on (a) patient and (b) slice level computed from the medians of the ROC curves of the test set predictions by our proposed method after each 30 training iterations

Fig. 6

Examples of 16 slices of the first test set that were consistently classified either correctly or incorrectly during the 6 repetitions of the first five-fold cross-validation split, including (a) positive slices correctly classified as positive, (b) positive slices incorrectly classified as negative, (c) negative slices incorrectly classified as positive, and (d) negative slices correctly classified as negative by our proposed method. All the slices are from different patients