Augmenting Interpretation of Chest Radiographs With Deep Learning Probability Maps.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Deep Learning for Healthcare

Background:

• Chest radiographs are crucial for diagnosing pneumonia.
• Deep learning offers potential to enhance radiograph interpretation.
• Semantic segmentation presents an alternative to binary classification for pneumonia detection.

Purpose of the Study:

• To explore the feasibility of a semantic segmentation deep learning model for identifying pneumonia foci on chest radiographs.
• To evaluate an alternative deep learning strategy beyond traditional binary classification.

Main Methods:

• A U-net convolutional neural network (CNN) was trained on 22,000 public chest radiographs with radiologist-defined bounding boxes.
• The model predicted pixel-wise probability maps for pneumonia.
• Performance was assessed using Dice overlap for localization and area under the receiver-operator characteristic curve for classification on an independent validation set.

Main Results:

• The model achieved an area under the receiver-operator characteristic curve of 0.854 for pneumonia classification, with 82.8% sensitivity and 72.6% specificity.
• Probability maps successfully localized pneumonia to the lung parenchyma in most validation cases.
• The mean Dice score for pneumonia segmentation in positive cases was 0.603, with 60% of cases having a Dice score >0.5.

Conclusions:

• Semantic segmentation using deep learning can generate probabilistic maps to aid pneumonia diagnosis.
• This AI-driven approach has the potential to expedite and improve diagnostic accuracy when used with clinical information.