Localizing Nodules in Chest Radiographs: Deep Learning-based Image Processing for Multi-Center, Multi-Device Validation
- Nov. 2020
- by Seo Taek Kong et. al.
To remedy the performance degradation of deep neural networks (DNNs) in identifying and localizing nodular patterns in chest radiography (CXR) apparent in multi-center, multi-device validation settings.
Materials and Methods
Five detection networks were ensembled after trained on an in-house dataset consisting of 4,436 nodular digital CXR images with lesion-level annotations collected from two major hospitals based in anonymized. An image processing network, whose outputs hereafter referred to as processed, was trained on the same data without using labels. This network processes test images in real-time with minimal perceptual distortion using a reference image selected from the training set.
The Japanese Society of Radiological Technology (JSRT) dataset, consisting of 154 nodular and 93 non-nodular CXR images acquired across 13 medical centers, was used to validate the proposed algorithm. Testing on this particular dataset resembles the multi-center and multi-device validation setting, as all source hospitals in the test set are geographically distant from the in-house data of origin, and images in the JSRT dataset are digitized chest films. The ensemble was tested on CXR images processed using the proposed algorithm and against two baselines: raw and manually filtered images using an in-house bandpass filter.
The ensemble achieved detection sensitivities of 48.7, 61.0, and 66.2 on raw, filtered, and processed respectively at false positives per image rate 1. A fine-grained analysis comparing the performances of all algorithms across different levels of difficulty show that processing especially aids in localizing subtle patterns previously overlooked by the detection network. The detection network’s sensitivity on processed images is comparable to general radiologists.
A style-transfer based image processing algorithm is proposed to process images at test time to better identify and localize nodular patterns in images collected from different medical centers or using unfamiliar devices. Unlike many nonlinear image processing techniques, the predictions on processed images can be easily interpreted on the original images.