Artificial Intelligence (AI) is an emerging technology in the field of medical imaging. The term AI is used to describe computer systems that mimic human cognitive functions, such as learning and problem solving. (1) It is the final achievement of a more complex technology, deep learning (DL), a subset of machine learning (2): these systems were built on flexible mathematical models known as deep neural networks (DNNs), a composition of multi-level mathematical algorithms able to identify complex non-linear relationships within large datasets. (3)
In particular, AI found fertile ground for development in the massive growth of data production and in the improvement of computing power. In that scenario, the exponential increase of digital radiological images, such as radiography, angiography, ultrasound, computed tomography (CT) and magnetic resonance imaging, makes radiological imaging one of the most prolific generators of digital data. (4)
In the past decade, and particularly in the last few years, AI-based technologies showed the potential to improve various steps of the radiological workflow and became essential “tools” in radiological imaging.
The most promising clinical applications of AI, currently reported in the literature, are:
- image analysis, segmentation and classification;
- denoising, image quality improvement and dose optimization;
- computer-aided diagnosis (CAD). (5)
With regard to CT, despite its indisputable role in the diagnosis and management of a large amount of major diseases, the potential risk of radiation-induced malignancy is still a primary weak point: reducing radiation dose exposure to “as low as reasonably achievable” (ALARA) is currently the main strategy for decreasing this risk. (6)
Thus, among the various tasks in which AI has shown promise, image denoising for CT radiation dose optimization is the one that raises the greatest expectations.
From Theory to Practice
The search for new tools for decreasing radiation doses without sacrificing image quality remains one of the main issues in the field of radiological imaging and radioprotection.
Deep learning-based reconstruction (DLR) protocols represent a practical implementation of AI in the daily radiological routine. DLRs are obtained through the application of deep convolutional neural networks (DCNNs), a type of machine learning subset widely applied for CAD: DLRs provide less noisy images than the standard reconstruction protocols at reduced radiation dose settings. The assumption is basically that with better reconstructions, high-quality ultra-low dose CT could be possible.
Based on these principles, Higaki et al. (7) compared the DLR technique to filtered back projection (FBP) and iterative reconstruction (IR). A 200 mm acrylic cylinder phantom was scanned with different tube currents, ranging from 10 to 100 mA with an interval of 10 mA, at the same tube voltage (120 kV); the images obtained were reconstructed with both IR (AIDR 3D, Canon Medical Systems Corporation) and DLR, and then the image quality and noise level were analyzed. The results demonstrated that the application of DLR allowed the reconstruction of less noisy images, despite low radiation doses, compared to FBP and IR techniques.
Another application of DCNNs is the residual neural network (ResNet), a specific type of neural network that provides a different learning approach to improve low-dose CT images. (8) The complex mathematical matrix behind this method can be simplified in a two-phase algorithm.
Firstly, in the network training phase, the algorithms analyze all complex information from two or more different samples of “input data”. During this phase, ResNet learns to recognize the differences, or “residuals”, among all input data samples analyzed: for example, if we consider the information from CT images obtained with two different radiation doses as input data, during the first phase ResNet records differences in terms of image quality or noise between each of the input data groups (e.g. group A 80 kV; group B 120 kV). Obviously, the more data are provided, the more powerful becomes the performance of the algorithms.
The “image postprocessing” phase is the second step of this DL algorithm: the well-trained network predicts the residual information between a new single input data and other previously analyzed data, to generate an “expected output data”, that should be as similar as possible to them. For example, if we consider the group A images as input data and the group B images as target data, ResNet should be able to subtract the residual noise pattern from group A to develop a “simulated group B image”.
These techniques were clinically applied by MacDougall (9) and by Cong et al. (10), who built ResNet algorithms to evaluate the performance of DCNNs applied to improve the image resolution of low-dose abdominal CT in pediatric patients and of few-view cone-beam breast CT, respectively. Both studies showed that a trained ResNet could be able to improve image quality, providing a significant reduction of the radiation dose delivered by a standard CT.
Conclusion
In the last thirty years, the world of radiological imaging has profoundly changed thanks to advances in the field of mathematics and informatics. Examples of the upheaval and renewal of radiological imaging can be observed in a number of areas, including the field of radioprotection and dose optimization: if compared to three decades ago, CT examinations show a 75% reduction of dose exposure. (11)
As already happened with IR applications, AI techniques will surely play a major role in the near future in improving image quality, reducing radiation dose and optimizing the workflow of CT examinations.
However, despite the promising preliminary results, further studies and investigations are needed and research in this field must continue to further improve AI technology. Finally, when prospective multicenter trials will demonstrate the safety and effectiveness of the clinical applications of AI, that is when radiological imaging will completely change.
References
- Russell S, Bohannon J. Artificial intelligence. Fears of an AI pioneer. Science. 2015 Jul 17;349(6245):252.
- Hosny A, Parmar C, Quackenbush J et al. Artificial intelligence in radiology. Nat Rev Cancer. 2018 Aug;18(8):500-510.
- Miller DD, Brown EW. Artificial intelligence in medical practice: The question to the answer? Am J Med. 2018 Feb;131(2):129-133.
- European Society of Radiology (ESR). Impact of artificial intelligence on radiology: a EuroAIM survey among members of the European Society of Radiology. Insights Imaging. 2019 Oct 31;10(1):105.
- Nichols JA, Herbert Chan HW, Baker MAB. Machine learning: applications of artificial intelligence to imaging and diagnosis. Biophys Rev. 2019 Feb;11(1):111-118.
- National Council on Radiation Protection and Measurements. Scientific Committee 46-3 on ALARA for Occupationally-Exposed Individuals in Clinical Radiology. Report No. 107 – Implementation of the principle of as low as reasonably achievable (ALARA) for medical and dental personnel: Recommendations of the National Council on Radiation Protection and Measurements. The Council; 1990.
- Higaki T, Nakamura Y, Zhou J et al. Deep learning reconstruction at CT: Phantom study of the image characteristics. Acad Radiol. 2020 Jan;27(1):82-87.
- He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770-778.
- MacDougall RD, Zhang Y, Callahan MJ et al. Improving low-dose pediatric abdominal CT by using convolutional neural networks. Radiol Artif Intell. 2019 Nov 27;1(6):e180087.
- Cong W, Shan H, Zhang X et al. Deep-learning-based breast CT for radiation dose reduction. In: Proceedings of the SPIE, Volume 11113, id. 111131L 7 pp. (2019).
- McCollough CH, Leng S. Use of artificial intelligence in computed tomography dose optimisation. Ann ICRP. 2020 Dec;49(1_suppl):113-125.