Application of neural networks to model the signal-dependent noise of a digital breast tomosynthesis unit

This work presents a practical method for estimating the spatially-varying gain of the signal-dependent portion of the noise from a digital breast tomosynthesis (DBT) system. A number of image processing algorithms require previous knowledge of the noise properties of a DBT unit. However, this information is not easily available and thus must be estimated. The estimation of such parameters requires a large number of calibration images, as it often changes with acquisition angle, spatial position and radiographic factors. This could represent a barrier in the algorithm’s deployment, mainly for clinical applications. Thus, we modeled the gain of the Poisson noise of a commercially available DBT unit as a function of the radiographic factors, acquisition angle, and pixel position. First, we measured the noise parameters of a clinical DBT unit by acquiring 36 sets of calibration images (raw projections) using uniform phantoms of different thicknesses, within a range of radiographic factors commonly used in clinical practice. With this information, we trained a multilayer perceptron artificial neural network (MLP-ANN) to predict the gain of the Poisson noise automatically as a function of the acquisition setup. Furthermore, we varied the number of calibration images in the learning step of the MLP-ANN to determine the minimum number of images necessary to obtain an accurate model. Results show that the MLP-ANN was able to yield the desired parameters with average error of less than 2%, using a learning dataset limited to only seven sets of calibration images. The accuracy of the model, along with its computational efficiency, makes this method an attractive tool for clinical image-based applications.