Purpose: While deep-neural-network (DNN) classifiers have arisen as promising tools showing state-of-the-art performances for detecting abnomalies in chest x-ray (CXR), one of the challenging problems that hampers the networks' performance is the data heterogeneity originated from a variety of sources such as devices, scan conditions, and image processing algorithms. We investigate the effect of a multi-frequency-based data normalization technique for the DNN-based diagnosis.
Methods: The reference data set consists of hundreds of normal and abnormal CXR images collected from a single hospital. Each CXR image has been decomposed into various frequency bands by use of the Gaussian-Laplacian pyramids technique. We calculated five levels of Laplacian pyramids. Then, five images at the Laplacian pyramids and the highest level image at the Gaussian pyramids were used for data normalization. We calculated the standard deviation value of each frequency band image. The average value of the Gaussian pyramid image was calculated as well. After that, we set the reference values at each frequency band by averaging calculated values. Each image in the reference database has been normalized iteratively to meet a convergence condition. We trained a DNN classifier with the normalized reference dataset. In the testing phase, we collected another CXR database that has different frequency characteristics. We normalized the test dataset in the same way.
Results: The inference of the non-reference dataset using the trained network combined with the data normalization resulted in an average precision score about 20% higher than that of the other trained network without data normalization.
Conclusion: We have developed a multiscale data normalization algorithm to increase the performance of the deep-neural-network classifier. Our study is believed to play an important role in developing many DNN CXR classifiers.