Daniele Della Latta PhD, Presenter: Nothing to Disclose

Marta Patronelli, Abstract Co-Author: Nothing to Disclose

Giovanna Letizia Di Girolamo BEng, Abstract Co-Author: Nothing to Disclose

Francesco D'errico, Abstract Co-Author: Nothing to Disclose

Dante Chiappino MD, Abstract Co-Author: Nothing to Disclose

We propose a model based on neural network to mimic radiologist perception of CT image. The BIQUNET was developed in Java and it is able to reproduce what a customer would  think about  image quality (IQ) during watching images .

Subjective and objective metrics of  IQ were used to train BIQUNET. We acquired image from a cylindrical acrylic phantoms and from patients (Chest) using Multi-Detector CT scanner (Aquilion 64,Toshiba Medical Systems, Japan). 9 different protocols in helical acquisition mode were used. The objective metrics of  IQ (contrast, noise and uniformity) were extracted from Dicom image file using an ad-hoc software. A group of 15 radiologist was asked to estimate the IQ derived from CT scanning. A classification of test images was realized analyzing observers scores in terms of perceived contrast and resolution. BIQUNET is a feed-forward neural network software architecture customized to allow supervised back-propagation training: it processes objective features extracted from CT images and it returns a quality score similiar to the one would expressed the radiologist. The architecture was chosen with test on phantom imaging based on Occam principle. A cross-validation approach measured the performance of the quality-assessment system.

The results from neural network performance are compared with the results from the human performance in terms of  perceived IQ. The neural network measures the quality of an image by predicting human observers score, using a set of objective features extracted from CT images. Experimental results show that the neural network outputs correlate highly (ρ=0.93) with observers scores, and, therefore, the neural network can easily be useful as subjective image quality assessment.

Human perception of image quality plays an important role in radiology. Obtaining subjective quality data is tedious and expensive. Using a back-propagation neural-network, we mimic radiologist assessment using objective IQ parameters  of CT images.

The goal in diagnostic imaging is to achieve diagnostically acceptable image quality at acceptable doses of radiation. BIQUNET could be oriented to looking for optimization of CT protocols.

Della Latta, D, Patronelli, M, Di Girolamo, G, D'errico, F, Chiappino, D, Computer Tomography Image Quality Assessment Using a Neural Network Java Approach: BIQUNET (Biomedical Image Quality NEural NETwork).  Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9014054.html