Introduction:
Computed tomography (CT) is widely used as an imaging technique in healthcare settings. On the other hand, because CT has such a high radiation dose, there are worries about possible radiation damage. Lowering the amount of X-ray radiation during CT scanning could lead to a notable deterioration in image quality, increasing the likelihood of incorrect and missing diagnoses. In the realm of CT, reducing the radiation dose and obtaining images of outstanding quality to satisfy medical diagnostic needs have long been important research priorities and challenges. Researchers have worked hard over the years to improve low-dose CT (LDCT) imaging algorithms, and deep learning (DL) -based algorithms (artificial intelligence) have proven to perform better than others.
What Are the Various Image Reconstruction Techniques?
Since the development of CT scanning, FBP (filtered back projection) techniques have been and continue to be the primary image reconstruction method on many modern scanners. FBP offers adequate image quality without requiring a significant amount of processing power. On the other hand, image noise and artifacts in FBP images can significantly rise with low radiation dose CT. I-RT was introduced into clinical practice in the early 2000s as a result of increases in processing power, which addressed the constraints of FBP at low radiation doses. I-RT (iterative reconstruction technique) has been the standard for image reconstruction for all CT protocols and examinations. However, I-RT showed changes in the appearance and texture of the images. DL-based image reconstruction is further introduced to overcome the drawback of I-Rt and further reduce the radiation dosage.
-
Filtered Back Projection: The conventional technique for reconstructing images in CT involves projecting multiple image data backward at the angle at which the images were taken. FBP-based image reconstruction selects sharp edges by generating negative image squares at the image margin, hence disregarding flat areas. This lessens fuzzy margins and improves image sharpness. Due to its speed and reasonable clarity of image, FBP continues to be the most used technique for image reconstruction. However, FBP approaches, which are linked to increased picture noise and artifacts, are no longer suitable due to the development of scanners that can produce thinner sections and the need to minimize the related radiation dosage.
-
Iterative Reconstruction Technique: The first technology for reconstructing CT pictures was IR, but due to its low processing power and lengthy processing times, it was not practical to use in clinical settings. However, recent advancements in technology have made this technique relevant again. It generates an array of estimated data, evaluated against measured data several times until the discrepancy between the two is as small as possible. However, IR algorithms have some drawbacks. A major problem is the lengthy processing time.
To further reduce dosage by improving image quality, artificial intelligence is a new approach being researched for CT image reconstruction. The general term for machines that display traits of human intellect is artificial intelligence (AI). AI algorithms in radiology are frequently trained to identify intricate patterns in the input data and generate numerical evaluations of particular imaging characteristics. Artificial intelligence can be attained in a variety of ways. One approach is through the use of machine learning (ML) and its category, deep learning (DL), in CT image reconstruction.
What Is Machine Learning in Image Reconstruction of CT?
A kind of artificial intelligence called machine learning uses vast amounts of properly labeled, structured data to identify patterns. The objective is to train machine learning models so that they can eventually predict information that is not labeled. The ML (machine learning) model identifies and links certain characteristics that are predefined and designed under human involvement to particular radiography findings. Supervised and unsupervised learning are two categories for machine learning.
Unsupervised learning refers to models that include no tags or characteristics specified under human supervision during the training process, whereas supervised learning calls for the model to be taught using labels or characteristics predetermined by humans.
Deep Learning Reconstruction:
Deep learning reconstruction (DLR) is a type of artificial intelligence-based CT reconstruction that offers a novel method of improving the quality of images without raising the radiation dosage to patients or the reconstruction time. Deep convolutional neural networks (DCNNs) are used in commercially available DLR techniques to identify and eliminate image noise patterns from either the raw data or the reconstructed computed tomography image.
Deep neural networks employ numerous layers of artificial neurons that may conduct computational calculations for a given problem in deep learning (DL), the dominant form of machine learning. Separate mathematical computations are used for measurable imaging and radionics, as well as machine learning, both supervised and unsupervised, to add diagnostic information. Beyond human capacity, DL can process large amounts of information, acquire knowledge, and use a variety of model frameworks. DL is capable of extracting significant features, that is, input for a particular job and data representations from enormous amounts of labeled and unlabeled data. The ability to learn data representations eliminates the requirement for hand-crafted features in DL, in contrast to standard ML. As a result, DL applications for object identification, picture segmentation, and categorization in radiography are growing. Although image analysis is the primary use of DL, which is very helpful in radiology, DL has found an additional application in CT image reconstruction.
Some have suggested replacing the current image reconstruction approaches with DL-based algorithms due to their encouraging finding. DL-based image reconstruction algorithms have been developed through a variety of methods, including image-based reconstruction that derives low-noise images with maintained resolution by using prior complex functions and CNNs trained on low-dose noisy CT images to produce pictures with the characteristics of routine standard dose acquisition.
Conclusion:
Since the initial CT image reconstruction technique was introduced, improvements in hardware and software have made it possible to minimize radiation doses, scan more quickly, and provide better image quality. Specifically, faster computing has made new possibilities possible. When compared to IR (iterative reconstruction) approaches, ML and DL inside AI promise greater image quality with additional dose reduction. As a developing technology, DLR, however, has not yet received enough evidence to justify widespread clinical use; just two DLR algorithms are currently authorized by the Food and Drug Administration and are available for commercial production.
