Towards appropriate use of test phantoms in training deep learning models for mammographic image conversion
Zahra Ghanian, Andreu Badal, Nicholas Petrick, Berkman Sahiner · 2023
We are developing machine learning methods for converting images from a source system to a target image acquisition system. The generated images preserve the anatomical content of a source image and produce the physical characteristics of a target image acquisition system. Our previous results indicated that incorporating slanted edge images into the training dataset improved the performance of the conditional generative adversarial network (cGAN) used for the image conversion. In this study, we investigated the appropriate use of phantom images in training when similar phantoms are used to measure the physical characteristics of the model-generated images. We performed two experiments with different training and testing data settings to study the bias that might be introduced when measuring the high contrast edge preservation (HCEP) capability of the cGAN with edge phantoms used as part of the training dataset. HCEP was measured using techniques borrowed from the measurement of the modulation transfer function (MTF). Our results indicated that using similar edge images in training the cGAN and then in measuring HCEP resulted in a biased estimate of the network performance in reproducing the target system MTF. Augmenting the training phantom data so that the training images were diverse and did not match the edge angle of the test phantom images reduced the bias and improved cGAN performance in modeling the target acquisition system. Moreover, this technique led to an improvement in the multi-scale structural similarity (MS-SSIM) value between the generated and the target breast images. Our results indicated that caution must be exercised when including phantom images in the training of a deep learning network if similar phantom images are also planned to be used for obtaining physical measures of image quality such as the HCEP or MTF.