A simple method for low-contrast detectability, image quality and dose optimisation with CT iterative reconstruction algorithms and model observers

The overall per caput mean effective dose per year to the population in European countries, due to X-ray procedures, is about 1.05 mSv. Computed tomography (CT), which is a key medical imaging modality within clinical diagnostic applications, contributes, on average, to 57% of this dose (range 5.31–83.1%) [1], with a mean value of 7.44 mSv [2]. In Switzerland, through 2013 the number of CT exams was 117 per 1000 inhabitants, with an average dose per exam of 8.54 mSv. CT alone contributed to about 70% of the collective dose, with an average annual effective dose of 1 mSv per inhabitant [3]. In light of these data, reduction of radiation dose from CT has become an essential field of study. In the last years, the advent of faster microprocessors, CT iterative reconstruction (IR) methods were launched to integrate already existing algorithms such as filtered back projection (FBP) as a way to reduce patient radiation exposure while maintaining high-contrast spatial resolution.

iDose4 (Philips iDose4™ system, Philips Healthcare, Cleveland, OH, USA) belongs to the first generation of iterative hybrid reconstruction algorithms, which combine FBP and IR algorithms [4]. On the contrary, Iterative Model Reconstruction (IMR; Philips Healthcare, Cleveland, OH, USA) is an advanced knowledge-based algorithm that models the process of physical data acquisition through the iterative minimisation of the differences between image raw data and the estimated image [5]. IMR differs from FBP methods in that the reconstruction becomes an optimisation process that takes into account data statistics, image statistics, and system models. IMR levels differ by number of processing cycles which increase concurrently with increasing levels. The main difficulty is to preserve an adequate diagnostic image quality reducing exposure mAs values and therefore reducing the dose to the patient [6, 7].

Medical image quality assessment involves both a scientific and philosophical approach to define how ‘well’ specific information of interest is obtained from images. One method to define medical image quality is called ‘statistical task-based assessment approach’ and consists of the evaluation of the observer performance during tasks such as patient classification or estimation of volume and/or other characteristics of tumours. However, studies based on human observers are resource-demanding and involve a significant variability of intra-observer and interobserver performance. Being able to extract as much statistical information as possible from the available images, computational model observers can be used as convenient and objective surrogates of human beings to predict and/or define their expected performance [6, 8].

In medical imaging, model observers were developed to study how system parameters affect signal detection [9], taking into account physical factors that degrade image quality. They are also useful to evaluate and optimise software systems, such as image reconstruction or processing methods, both to study and predict their effects on human-observer performance [10–12].

The purpose of this work was to evaluate image quality and low-contrast object detectability in CT phantom images acquired at different tube loadings (i.e. mAs) and reconstructed with different algorithms in order to reduce mAs and consequently CT dose, with respect to a standard reference value, without detriment of the images. Model observer performance in terms of minimum diameter of the detectable low-contrast details was also evaluated.

The purpose of this work was to use IR algorithms for obtaining a percentage threshold value of mAs in order to reduce CT dose while maintaining image quality. Human and computational detection performances were also evaluated. In general, results obtained by means of CTQA_cp and Catphan QA in terms of image quality were approximately in agreement. The resolution of a CT imaging system is well characterised with the MTF, which indicates its ability to reproduce various levels of detail from a region of the patient to its image. Small differences obtained for uniformity and MTF are likely due to small differences between the applied calculation algorithms. In particular, for uniformity analysis, position and dimension of the ROIs may change between CTQA_cp and Catphan QA. It is only specified that in CTQA_cp the area of the ROIs correspond to the area of a circle with diameter 10% of the diameter of the homogeneous region in the CP486 module. Whereas for Catphan QA, it is indicated that the outer edge of each ROI is located 1 cm from module border. For MTF evaluation, the small changes might be due to differences in the Fourier transform analysis of the images.

Image quality analysis anyway confirmed data already reported in literature, supporting the efficiency of the novel IR methods if compared to standard reconstruction algorithms such as FBP [17, 18]. Regarding noise, it initially increased while mAs values were lowered using FBP reconstruction. The application of iDose constantly reduced it, even at decreasing mAs, and IMR kept it low while tube loading reduced to 50 mAs. Below this value, noise increased with a consequent degradation of the image quality due to IR limits at very low mAs values [19]. It was observed that the uniformity values were within the advised limit (ΔHU ≤ 4) [20]. iDOSE provided, therefore, a similar image resolution to that obtained with FBP at significantly higher mAs values. Our results represent, therefore, a valuable confirmation that the use of IR algorithms preserves the spatial resolution while reducing mAs [21], except at very low tube load (i.e. < 50 mAs) where a very small decrease in spatial resolution was found [19].

Evaluation of the low-contrast performance in CT imaging is a difficult task. It is related to the ability of an operator to distinguish between two objects or regions with similar CT number and it depends on statistical noise levels, contrast and size of the signal.

Referring to Fig. 5, on the one hand the percentage of correct answers is a proper quantification of the efficiency of the application of the various reconstruction algorithms for low-contrast details identification, on the other hand the standard deviation is a good descriptor of inter-observer’s variability of image quality evaluation. The comparison between the different acquisition and image reconstruction modalities confirmed the highest efficiency for IMR, level 3 [22]. In fact, in all tested conditions low-contrast detection rates were greater for IMR than for FBP or iDOSE; low-contrast detectability was preserved with IMR up to a tube loading reduction to 40 mAs. In accordance to Katsura et al. [23], a consequent decrease of dose to the patient by a factor up to 7 (i.e. 80% dose reduction) seems, therefore, to be possible without producing any significant detriment to the images.

Interestingly, the variability of the percentage of correct answers was high for iDose in the range of 220–160 mAs, whereas it was much lower with IMR, even at reduced tube loading. This was due to possible reconstruction limitations of iDOSE, which might stress the perception variability among observers. As previously described, psychological factors could, in fact, affect the test results [10]. As observers have performed the test in different moments of the day, diagnostic accuracy, visual accommodation, reading time, subjective ratings of fatigue and visual strain, before and after a day of clinical reading, may all have contributed as confounding factors in terms of image quality evaluations. Image texture, artefacts and over-smoothing of images with higher strengths of IR may have affected diagnostic results [24].

One could argue that, different from our work, a large amount of papers compared iterative with exact methods for specific clinical applications (e.g. morphological evaluation of tumours) in specific anatomical regions and diseases. We did not focus on specific anatomical regions and diseases because the approach described in this study could be adopted in many different clinical applications, including the low-contrast regions/tissues detection task, such as liver lesion identification in abdominal CT [24, 25].

The performance of model observer software was, in general, good in terms of low-contrast detectability up to 100 mAs. Below this value, the model observer did not work well probably due to software intrinsic limitations. Different from human observers, model observer software recognised objects of 2 mm in diameter as a prediction of human observer performance also between 220 and 160 mAs. This difference was particularly evident on the 220-mAs images, where the mean value of correct answers by human observers was 9%, whereas according to the model observer a 2-mm diameter object is detectable.

The model observer given in CTQA_cp showed to be a valid tool for a first evaluation of the analysed data, but presented the following limitations that would require upgrades and improvements: (1) no optimisation/adaptation is possible to ‘instruct’ the system for specific study conditions; (2) no univocal and absolute detectability scoring is provided as output. In particular, a detectability scoring could be important to better quantify the right mAs reduction percentage, optimised in terms of human-perceived image quality. In general, a better model observer software, with sophisticated interfaces and specific setup possibilities, should probably be adopted in future to assist better and predict human observer’s performance. Implementation of a more advanced software, which is beyond the aim of this study, is, however, very complex, as it requires a thorough knowledge of model observer theory, statistics and informatics [26].

In conclusion, this study demonstrated that the application of the IR algorithm IMR to phantom images preserves a good image quality and object detectability for human radiological evaluation of CT exams, with a potential noise reduction up to 60% and, in particular, an 85% dose reduction to the patient. With respect to other studies, the method presented in this work can be easily implemented and contains a thorough analysis for the evaluation and optimisation of mAs according to the adopted reconstruction algorithms. The model observer can, in principle, be useful to assist human performance in CT low-contrast detection tasks and in dose optimisation, but needs to be optimised in order to extract useful information to support and predict human observer evaluations on CT images. Further studies are required to confirm the reported findings.