KEYWORDS: Digital breast tomosynthesis, Image quality, Denoising, Human subjects, Reconstruction algorithms, Medical image reconstruction, Deep learning, 3D image reconstruction, Convolutional neural networks
Image noise in digital breast tomosynthesis (DBT) reduces the detectability of subtle signs of breast cancer such as microcalcifications (MC). This study investigated the potential of applying DNGAN, our previously developed deep convolutional neural network (CNN) DBT denoiser, to different reconstruction stages to improve the image quality, including projection views before reconstruction, intermediate images during iterative reconstruction, or final reconstructed images, and a combination of the different stages of denoising. We also proposed two CNNs as task-based image quality measures to compare different reconstructions: a CNN noise estimator (CNN-NE) trained to evaluate the noise level of a given DBT image, and a CNN MC classifier (CNN-MC) trained to estimate the detectability of MCs by classifying clustered MCs from MC-free backgrounds. The CNN-NE was trained with virtual DBTs reconstructed from projections generated by the VICTRE tool over a wide range of noise levels. The CNN-MC was trained with human subject DBTs. We adopted the training strategy of transfer learning to train CNN-NE and CNN-MC due to the limited training data. We found that the increase in AUC estimated by the CNN-MC classifier correlated well with the decrease in image noise by DNGAN estimated by CNN-NE on an independent human subject test set. A combination of DNGAN-regularized plug-and-play reconstruction and an additional DNGAN post-reconstruction denoising achieved the lowest noise level and the best MC detectability. The AUC and noise rankings from the CNNs matched our visual judgement that less noisy images had better MC conspicuity.
KEYWORDS: Breast imaging, Medical image reconstruction, Digital breast tomosynthesis, Reconstruction algorithms, Human subjects, Convolutional neural networks, Image quality, Denoising, Image restoration
Digital breast tomosynthesis (DBT) reconstruction is an ill-posed inverse problem due to the limited-angle acquisition geometry. DBT is also a low dose imaging technique and has very noisy projection views. In this study, we investigated the feasibility of improving image quality of DBT reconstruction by combining (1) a model-based iterative reconstruction (MBIR) method that models the detector blur and correlated noise (DBCN) of the DBT system, and (2) a deep convolutional neural network based DBT denoiser, DNGAN, that we developed in our previous work. DBCN is physics-based whereas DNGAN is data-driven. We followed the regularization by denoising (RED) framework to construct a regularizer from DNGAN and used the DBCN-modeled terms in the MBIR formulation. We solved the optimization problem using the proximal gradient method. The proposed approach, named DBCN+DNGAN, was tested on a set of human subject DBT data sets. The image quality was evaluated quantitatively with figures of merit (FOMs) including the contrast-to-noise ratio, full width at half maximum, and task-based detectability index of a set of microcalcifications individually marked in the human subject data set. We found that these FOMs were improved in the DBCN+DNGAN-reconstructed DBT volumes compared to those reconstructed with DBCN alone or with the simultaneous algebraic reconstruction technique. The soft tissue appearance was visually satisfactory and the background noise level was low in the DBCN+DNGAN reconstructed images.
KEYWORDS: Digital breast tomosynthesis, Denoising, Convolutional neural networks, X-ray imaging, Visualization, Human subjects, Data acquisition, Breast
This paper investigates the training of deep convolutional neural networks (DCNNs) to denoise digital breast tomosynthesis (DBT) images. In our approach, the DCNN was trained with high dose (HD)/low dose (LD) image pairs, in which the HD image was used as reference to guide the DCNN to learn to reduce noise of a corresponding input LD image. In the current work, we studied the effect of the dose level of the HD target images on the effectiveness of the trained denoiser. We generated a set of simulated DBT data with different dose levels using CatSim, a dedicated x-ray imaging simulator, in combination with virtual anthropomorphic breast phantoms simulated by VICTRE. We found that a DCNN trained with higher dose level of HD target images led to less noisy denoised images. We also acquired DBT images of real physical phantoms containing simulated microcalcifications (MCs) for training and validation. The denoisers trained with either simulated or physical phantom data improved significantly (𝑝𝑝 < 0.0001) the contrast-tonoise ratio of MCs in the validation phantom images. In an independent test set of human subject DBTs, the MCs became more conspicuous, and the mass margins and spiculations were well preserved. The study showed that the denoising DCNN was robust in that the denoiser could be trained with either simulation or physical phantom data. Moreover, the denoiser trained with CatSim simulation data was directly applicable to human DBTs, allowing flexibility in terms of the training data preparation, especially the HD images.
KEYWORDS: Digital breast tomosynthesis, Denoising, Convolutional neural networks, Image processing, Image denoising, Breast imaging, 3D image reconstruction
To reduce noise and enhance the contrast-to-noise ratio (CNR) of microcalcifications (MCs) in digital breast tomosynthesis (DBT), we conducted a study to investigate the feasibility of performing denoising on the projection views (PVs) using a deep convolutional neural network (DCNN) before reconstruction. We fine-tuned a modularized adaptive processing neural network (MAPNN) based on a pretrained model for CT image denoising. Four phantom DBTs containing over 700 simulated MCs of 3 nominal sizes were scanned with a DBT system that acquired 9 PVs within a 25° scan angle at two dose level settings to form the training and validation sets. Nine human subject DBTs were used as an independent test set. We trained the DCNN with low dose PVs as the input and the corresponding high dose PVs as the reference. We marked the MCs in the PVs and designed a loss function for DCNN training that balanced the effect of noise reduction and signal preservation. The loss function was a weighted sum of the perceptual loss, the adversarial loss and the CNR loss. A visual comparison of the DBT volumes reconstructed from the denoised PVs indicated that the proposed method could reduce noise and preserve the texture of the background without blurring subtle MC signals. A quantitative comparison showed a significant CNR improvement (p < 0.0001) in both the validation phantom and the human subject DBTs.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.