Introduction
Over the past decade, software developers have contributed to an extreme technology leap. It contributed to information extraction from images and popularized the notion of convolutional neural networks. According to Lei et al. (2019), such neural networks can be significantly more productive than their outdated alternatives because they process image data more quickly. The ability to step away from classical machine learning techniques could be one of the main reasons why modern image processing was improved, with numerous features being added to the pool.
Deep learning methods turned out to be an opportunity for medical professionals to gain new insights and learn new technology-related skills (Boulanger et al., 2021). The topic of the current literature review is an exploration of magnetic resonance (MR)–based synthetic computer tomography (CT) generation using deep learning methods for the purposes of radiotherapy in areas of the brain, head, and neck.
Results
Convolutional Neural Networks
When training a CNN, researchers may expect the imaging tools to help convert a conventional MR scan into a synthesized CT image. This is one of the main reasons why Dinkla et al. (2018) tried to find a balance between dedicated MR sequences and required calculations. The need for a generic solution made them understand that CNNs could function without the acquisition of additional sequences to enhance the signal. Despite the existence of MR-only frameworks, experiments were continued to see how minimum scan times could enhance the outcomes of the whole process.
Accordingly, the new workflow would require recalculations and altered dose calculations. Another article by Dinkla et al. (2019) made it safe to say that patient selection did not have to be limited to help researchers find the most valid approaches to the implementation of CNN. It was proposed that attention be paid to small bony structures and pathology-induced differences be observed to enhance the clinical relevance of sCTs. Atypical MRI appearances were no longer seen as a display of uncertainty and negligible dose calculation.
An updated variation of classical neural networks can be powered by deep learning technologies. With a deep convolutional neural network (DCNN), image generation and processing could become even easier, paving the way for simpler model deployment (Han, 2017). One major advantage of this framework would be enhanced computation time. The amount of time required to train the system would increase significantly, but it would then lead to the creation of sCTs that can run on single-GPU systems. Therefore, DCNN is a potential shortcut that can advance modern models of image processing.
According to Han (2017), patch fusion might accelerate the process of innovation even more. At the same time, the decision to reduce computation time could lead to an increased occurrence of errors affecting the quality of the generated image. At the moment, there are no recommendations regarding potential improvements. Regardless, DCNN is an undeniable contributor to the development of high-quality imaging and deep learning as a whole.
Deep Learning
The second venue of research that has to be reviewed within the framework of the current literature review is the influence of deep learning on the quality of imaging procedures. Boulanger et al. (2021) suggest that the advent of deep learning became one of the main reasons why radiotherapy began benefiting from high-quality CT and MRI.
The variety of applications of deep learning makes it safe to say that there could be numerous digital sequences affecting the process depending on the necessary functions and anatomical locations. Deep learning also gave rise to quite a few metrics that can be utilized to evaluate the efficiency of sCT, such as geometric fidelity and voxel intensity (Boulanger et al., 2021). With deep learning, the process of head localization became significantly easier, especially considering multiple inputs and their impact on the quality of sCT.
These findings can be supported by the fact that deep learning tools can be trained respectively, expanding the horizons related to sequence types, fundamental calculations, and image generation. Zimmermann et al. (2022) found that the general concept of deep learning can be expected to support the most recent developments in the area and get exposed to corrupted imagery less often. The learning rate of such tools can be aligned against normalization, especially with new features in place.
From ResNet blocks to U-net tools, numerous standardization items could be intended to perpetuate the importance of deep learning. According to Zimmermann et al. (2022), the outcomes of respective efforts would have to be seen as dependent on pretraining and the search for new features. Hence, deep learning requires a vast array of resources and several knowledgeable staff members.
Generative Adversarial Networks
The transition from MRI to CT has to be supported by additional tools in order to be of higher quality and contain vital information about the patient’s head and neck. Experiments carried out by Liu et al. (2021) prove that multi-cycle GAN could be one of the ways to achieve superior performance and gain access to extremely detailed images.
From mean absolute error to peak signal-to-noise ratio, multi-cycle GAN can bypass any other alternative model in terms of performance because of an incredible level of steadiness. The latter can be achieved through the interface of the Pseudo-Cycle Consistent module that mediates synthetic stability and enhances intermediate outputs. In a sense, applying multi-cycle GANs can highlight the difference between synthetic and natural CTs.
On the other hand, there are guarantees revolving around the idea that the output results of GANs can remain consistent regardless of the generator input data. Hence, the presence of an additional Domain Control module could enhance image correctness and make the generator much more stable overall (Yang et al., 2020). With this information at hand, responsible stakeholders could replace GAN generators proactively to achieve better results. More specifically, Z-Net could be the best option to improve imaging performance through the interface of skip connections. Hence, one of the key benefits of GANs is the possibility of combining high- and low-level features without compromising the ultimate quality of the image. With deeper networks, it could highlight unseen features and contribute to high-quality synthesis that is accurate and relevant.
Another important portion of evidence relates to the fact that there are similar network structures possessed by cGAN and U-net generators, making it safe to say that GAN methodology is rather precise. With limited data loss, it can be claimed that cGAN is so powerful because of its additional discriminator that can identify the realness of the input image (Qi et al., 2020).
A recent study by Qi et al. (2022) also suggests that the results above could be propagated back to the generator to update the overall parameters of the tool. GAN-based instruments can be expected to prevent image corruption and generate sCT images that are much closer to actual CT images than any of the previous alternatives. Even the amount of blur in the image can be mediated effectively, making the outcomes better than the U-net counterpart.
Notwithstanding the benefits of the GAN models described above, several limitations have been identified as well. For instance, Emami et al. (2018) claimed that such models could be prone to overgeneralization, especially when images are of poor quality. It could lead to mode collapse and inappropriate dataset processing, generating images that are not qualitative enough to contribute to enhanced care.
For some of the available systems, the GAN methodology could be too resource-intensive due to the need to resolve real-world issues and the overall lack of case evidence (Jabbarpour et al., 2022). Modern imaging modalities, such as the CycleGAN, have to be researched more profoundly to prepare updated datasets and enhance the quality of imaging. Knowing that pixel-wise alignment is not going to be necessary, the process will be quicker, and fewer errors or uncertainties will affect the outcome.
MRI-Only
One particular factor that has to be mentioned when discussing the effectiveness of MRI-only approaches to scanning and imaging is the generation of electron density maps and enhanced geometric accuracy. For example, Kazemifar et al. (2019) dwelled on how MRI-only benefits dose recalculation and the overall quality of sCT, with a significantly higher passing rate. Further treatment delivery also can be established on the basis of the results of MRI scans, which affects the quality of final verdicts.
The primary target for medical professionals utilizing an MRI-only approach would be to make all required calculations in advance and predict necessary dose levels to prevent poor imaging and patient outcomes. The accuracy of sCT can be deemed exceptionally high when treatment plans are clinically validated and sufficiently supported by evidence.
A significant benefit of MRI-only approaches was identified in another recent research project where robust reprocessing methods were tested. Lei et al. (2019) found that geometric artifact correction could be attained with the help of one of the simplest approaches to imaging. Examples are shown throughout their investigation to hint at the fact that geometric artifacts make the quality of imaging susceptible to the lack of detail. Nevertheless, new MRI-only scans can bypass geometric distortions with the help of altered magnitude of susceptibility effect, as field inhomogeneity mapping can be utilized to establish in vivo quality control (Lei et al., 2019). Most importantly, MRI-only imaging has to be reckoned with because it corrects the worst cases of geometric distortion and removes artifacts. Future research could focus on field map-based corrections powered by MRI results.
Discussion
The image evaluation process has to begin with a detailed outlook on how CT generation could benefit from the increasing number of alternatives. Deep learning methods hold a certain level of potential that requires training on natural images in order to avoid unnecessary bias. Such pretraining could become an opportunity to avoid discontinuities across images and make the model more robust (Dinkla et al., 2018; Dinkla et al., 2019). Interscan differences become less of a problem nowadays because the overall level of sensitivity of respective instruments can be modified as well. Machine-led calculations demonstrate a high level of accuracy while also helping generate images quicker from a single MR scan.
Another point to be addressed is the importance of DCNN for the accommodation of training data. The high capacity of deep learning models has to be tested in practice to generate even more data for further exploration (Han, 2017). In other words, training is an essential element of medical imaging because the number of variations of deep learning-based methods continues to increase on a daily basis. The accuracy of new imaging is high enough to prevent the degradation of data even in the long run.
The current literature review serves as proof of the fact that GAN models are extremely powerful and provide medical staff with accurate sequences and high-quality images of the neck and head region. An essential finding here is that Boulanger et al. (2021) suggest that CT predictions can be enhanced with the aid of GAN models. Even if MRI-only radiotherapy is going to be conducted in the future, it will create a paradigm where the accuracy of sCT generation is validated successfully.
Zimmermann et al. (2022) reviewed necessary image evaluation metrics and presupposed that quality assurance would still have to be implemented. It means that external radiotherapy treatments can be planned by multidisciplinary units that possess the required knowledge and training to implement digital solutions. When conducting MRI scanning, medical specialists will have to focus on data acquisition standards to see how sCTand GAN can enhance diagnosis and decision-making.
One more vital topic of discussion is the need to set up neural networks and conduct quality audits to investigate sCTgenerators. Innovative technology should not break clinical workflows, which is expected to lead to positive outcomes. As a result, comparisons conducted by Liu et al. (2021) and Yang et al. (2020) suggest that real-world data has to be acquired by deep learning systems to increase the complexity of medical imaging. Existing literature makes it safe to say that image translation is heavily powered by GAN and synthetic CT.
According to Qi et al. (2020) and Qi et al. (2022), some of the images generated with the aid of the former could be blurry. When transferring MRI to sCT, medical professionals may have to pay more attention to the quality of the original image and how detail conversion can enhance the final result. Additional experiments are required to develop GANs that can process dose calculation and multi-sequence images. The best performance could be achieved with complementary inputs in place, supporting MRI imaging and sCT.
Another insight that has to be perceived through the lens of necessary improvements is the increasing number of image-generation methods that lead to the development of state-of-the-art technologies. According to Jabbarpour et al. (2022), for example, MRI-only radiotherapy is not detailed enough to emerge as the best treatment planning support element. All the promising results have to be reviewed carefully to explain how image synthesis becomes more consistent over time, contributing to the removal of constraints between pre- and post-processed images.
Another idea is that anatomical geometry can be altered within the framework of synthetic CT, which also means that undesirable outcomes are still a possibility. Emami et al. (2018) suggested that the spectral normalization technique should be deployed more often to stabilize self-training procedures and make the best use of the self-attention module. Accurate and visually detailed images can be expected to generate improved medical services.
Additionally, it is vital to look at how GAN models are going to develop over time, as these methods can improve the whole clinical workflow and enhance the overall precision of operations. Kazemifar et al. (2019) suggested paying more attention to computational time and the possibility of going beyond outdated clinical strategies that only focused on synthetic CT images. Reduced training time is also essential when carrying out brain, head, and neck imaging. Even if some of the data remains unregistered, the new approach is extremely beneficial because it will only require medical professionals to streamline the process without worrying about CT and MRI images and their overall quality.
For new deep learning models, there is a major challenge: the lack of tools that distinguish between bone regions and air. This is why Lei et al. (2019) propose a renovated 3D GAN model intended to reduce calculation error. Such a substantial contribution to the quality and quantity of MRI and CT could affect treatment planning procedures in the future, requiring vast training datasets and high-quality synthetic images.
With MRI-only solutions, it is going to be possible to reduce the costs associated with clinical efficiency and radiation dose persevered by the patient (Emami et al., 2018). High-precision planning can become another advantage for medical professionals looking to develop and test new imaging strategies. At the moment, the most popular proposal is to generate more synthetic CT images to enhance radiation therapy and acquire high-quality outlooks on the given patient’s brain, head, and neck.
Conclusion
Overall, it has to be concluded that radiology significantly benefits from the implementation of deep learning methods. Various approaches to radiotherapy seem to transpire in state-of-the-art technology, and a number of promising interventions cannot be overlooked if researchers expect to benefit from deep learning as much as possible. Hence, the current literature review indicates that various branches of radiology can successfully implement improved image processing to acquire additional decision-making and diagnosis support. Apart from improved image processing, deep neural networks might save time and money by reconstructing images more quickly. Even if there are more than a few experimental approaches, the idea is that deep learning will dominate the MRT image processing field. Following the trends and accelerating the workflow is vital by deploying more relevant technologies when possible.
On the other hand, such disruption cannot be without limitations. For example, technical difficulties associated with image processing can lead to an imbalance between how multiple sequences are handled and what normalization techniques are applied. Deep learning models are somewhat underdeveloped in generating consistent results because there is no significant quantity of data intended to enhance medical imaging. The existence of data does not predict successful acquisition.
Hence, radiologists’ predictions will still prevail as the primary source of bias. Despite innovation, deep learning models will be affected by random variability and systemic training flaws. It is evident that deep learning systems exceed human performance by far, so new developments in the field should be stimulated to develop new GAN models and improve CT imaging.
Reference List
Boulanger, M. et al. (2021) ‘Deep learning methods to generate synthetic CT from MRI in radiotherapy: a literature review’, Physica Medica, 89, 265-281.
Dinkla, A. M. et al. (2018) ‘MR-only brain radiation therapy: dosimetric evaluation of synthetic CTs generated by a dilated convolutional neural network’, International Journal of Radiation Oncology* Biology* Physics, 102(4), 801-812.
Dinkla, A. M. et al. (2019) ‘Dosimetric evaluation of synthetic CT for head and neck radiotherapy generated by a patch‐based three‐dimensional convolutional neural network’, Medical Physics, 46(9), 4095-4104.
Emami, H. et al. (2018) ‘Generating synthetic CTs from magnetic resonance images using generative adversarial networks’, Medical Physics, 45(8), 3627-3636.
Han, X. (2017) ‘MR‐based synthetic CT generation using a deep convolutional neural network method’, Medical Physics, 44(4), 1408-1419.
Jabbarpour, A. et al. (2022) ‘Unsupervised pseudo-CTgeneration using heterogenous multicentric CT/MR images and CycleGAN: dosimetric assessment for 3D conformal radiotherapy’, Computers in Biology and Medicine, 143, 105277.
Kazemifar, S. et al. (2019) ‘MRI-only brain radiotherapy: assessing the dosimetric accuracy of synthetic CT images generated using a deep learning approach’, Radiotherapy and Oncology, 136, 56-63.
Lei, Y. et al. (2019) ‘MRI‐only based synthetic CT generation using dense cycle consistent generative adversarial networks’, Medical Physics, 46(8), 3565-3581.
Liu, Y. et al. (2021) ‘CT synthesis from MRI using multi-cycle GAN for head-and-neck radiation therapy’, Computerized Medical Imaging and Graphics, 91, 101953.
Qi, M. et al. (2020) ‘Multi‐sequence MR image‐based synthetic CT generation using a generative adversarial network for head and neck MRI‐only radiotherapy’, Medical Physics, 47(4), 1880-1894.
Qi, M. et al. (2022) ‘Multisequence MR‐generated sCT is promising for HNC MR‐only RT: a comprehensive evaluation of previously developed sCT generation networks’, Medical Physics, 49(4), 2150-2158.
Yang, H. et al. (2020) ‘Unsupervised MR-to-CT synthesis using structure-constrained CycleGAN’, IEEE Transactions on Medical Imaging, 39(12), 4249-4261.
Zimmermann, L. et al. (2022) ‘An MRI sequence independent convolutional neural network for synthetic head CT generation in proton therapy’, Zeitschrift für Medizinische Physik, 32(2), 218-227.