First-generation IR algorithms were mainly focused on simple noise reduction; radiologists complained about a blotchy image appearance. Second-generation algorithms are generally of the so-called statistical type. These algorithms take the statistical noise distribution into account. They typically exhibit fewer IR artifacts, whilst allowing increased noise reduction. 

Model or knowledge-based IR algorithms represent the third generation. They incorporate detailed physics knowledge of photon-target interaction, x-ray tube emission, detector element cross-talking, into the image reconstruction process, which permits much more accurate and patient-specific Monte-Carlo simulations. Thanks to highly increased and cheap computing power, recent high-end CT scanners can perform these calculations in real-time for routine exams

Pursuing a CT system's optimisation in accordance to the ALARA principle represents a challenge for medical physicists when IR algorithms come into play.

In presence of heterogeneity, noise properties of IR reconstructed images are highly non-stationary. Hence all well-known Fourier metrics as modulation transfer function (MTF) or noise power spectrum (NPS) should be applied cautiously to IR reconstructed images. 

Recently, Deep Learning Reconstruction Algorithms have been introduced in Clinical Practice, aiming to improve image quality, removing all the critical aspects related to the use of IR reconstruction algorithms.

This work aimed to investigate the mode of action of a Deep learning Reconstruction algorithm in comparison with 5 different commercial  IR algorithms and FBP of three different vendors by means of conventional image quality metrics.