April 29, 2022
Computed tomography (CT) facilitates early diagnosis of a wide range of diseases. When carrying out CT, a diagnostician faces a challenging task of ensuring that the examination is efficient and safe. During the examination, a patient receives an X-rat dose. But its reduction can lead to inadequate quality of the scan.
Methods of CT scanning and interpreting the results can be revised due to machine learning. Technologies based on artificial intelligence (AI) make it possible to strike a balance between image quality and radiation dose.
Advanced technologies expand the usual range of clinical tasks that the physician can solve using CT, from differential diagnosis to intelligent forecasting.
Machine learning algorithms make it possible to change the approach to the use of computed tomography:
to raise the safety level of the method. The artificial intelligence increases the value of the diagnostic procedure, since it allows for gaining more information from the received scans. A physician can, if necessary, carry out a reconstruction of an image from several slices and raise its quality by processing and filtering it. This does not require making a new CT scan. Computing processing replaces partially contrast medium, facilitating visualization of pathological areas. The examination can be carried out with lower radiation doses and contrast medium load1.
to detect subtle pathologies. Machine learning algorithms perform image segmentation with the identification of areas of interest. Identification of small-sized findings typical of the early stage of a disease is an important part of early diagnosis. The AI capabilities help a physician in this task. The CT Lungs service by SberMedAI can detect minimal nodules in lung tissue exceeding the limit of 4 millimeters2.
to reduce the time spent to interpret the image. The AI can automate routine processes, such as assessing the extent of damage, preparing a preliminary report, providing prompt exchange of information between physicians. Intellectual sorting of patients help the diagnostician to pay attention to severe cases first3.
to raise clinical reliability of the results. Machine learning algorithms make it possible to achieve high prediction rate and accuracy. They identify hidden patterns when analyzing the image and carry out a quantitative assessment of pathological area4. Visualization data can be combined with clinical indicators to predict the outcome of a disease5.
to use the resources of a medical institution rationally. The use of intellectual technologies influences the CT economic efficiency. It raises patient throughput, helps to achieve a compromise between radiation dose and image quality, and reduces wear and tear on medical equipment6.
Main causes of classification problems, which we may encounter when applying AI, include the following7,8:
datasets may contain images with different aspect ratios and color patterns, since CT scans may be taken on different machines and by different operators;
the discrepancy between the dataset, on which the intellectual system was developed, and the dataset, on which it was deployed;
presence of the image artifacts, which can be mistaken by the computer for pathology;
the use of classifiers of the ‘black box’ type: it is difficult to tell how the artificial intelligence came to this or that conclusion;
the choice of data architecture can affect the systematic classification error to varying degrees.
The appearing of the bias of the machine learning algorithm leads to an increased frequency of false-positive and false-negative results. Thus, the disease may be erroneously identified with a healthy person or not identified with a sick one.
The developers carefully select data for AI training, making high demands on the quality of images and their markup. Medical visualization data can be post-processed. Neural networks carry out image reconstruction, artefact elimination and denoising9.
The choice of explicable classification models and the comparison of their performance facilitate the reduction of systematic error and ‘black box’ effect. The latter helps the physician to understand the computing logic of the machine and make a correct diagnosis.
During the pandemic, computed tomography accelerated the process of viral pneumonia diagnosis in an environment with a high physician workload. Machine learning distinguish between respiratory diseases having different nature but similar symptoms10.
A research group from the US published the results of the machine learning algorithm development in Nature Communications. The algorithm can identify signs of COVID-19 by chest CT scans. Two classification models were tested as part of the research11:
a full 3D model used the entire lung area, transforming the image according to the preset size;
a hybrid 3D model created an image on the basis of several tomography slices.
To train, check and test the model, 2,724 scans of 2,617 patients were used, including those with confirmed COVID-19. The images used to train the model were preliminarily annotated by expert radiologists. The assessment of system efficiency was carried out in four independent institutions of China, Italy and Japan11.
In one of the tests, the researchers focused on pulmonary pathology diagnosis, excluding the consideration of extrapulmonary chest pathologies. A complete 3D model identified signs of viral COVID-19 pneumonia with 87 out of 109 patients. The hybrid model managed to identify the changes typical of the disease on the CT for 74 out of 109 patients11.
The researchers note that the algorithm had difficulties identifying the nature of a respiratory infection. They also pointed out that there is a probability of false positive result for pneumonias of bacterial, fungal, and viral nature not related to COVID-1911.
At the same time, the algorithm proved to be more accurate in the case of early pneumonia rather than a progressive course. Its application includes quantitative assessment of damaged tissue and characteristics of its changes11.
Scientists from China were interested in the problem of differential diagnosis in COVID-19 using artificial intelligence. The neural network COVNet is a development designed to efficiently distinguish signs of COVID-19 infection and community-acquired pneumonia on diagnostic scans12.
The working principle of the three-dimensional machine learning program was presented in the Radiology journal12:
The neural network uses CT scan fragments as input data;
it generates a function for the corresponding fragments;
the extracted elements are combined with the function map creation;
the probability is assessed for all pathologies: COVID-19 pneumonia, community acquired pneumonia and a disease that is not pneumonia.
To visualize damaged areas, the scientists introduces heat maps. For each forecasted class, the algorithm marks suspicious area in red, so it does not require additional manual annotation. This allows the radiologist to focus their attention when interpreting the examination results12.
COVNet showed good results in identifying COVID-1912:
sensitivity of the neural network was 90%, and specificity was 96%;
the neural network managed to distinguish between COVID-19 viral pneumonia and community acquired pneumonia with high accuracy: AUC classification indicator was 0.96 and 0.95 respectively.
Scientists from the Fudan University in Shanghai and Wuhan University considered the use of machine learning in CT scan analysis for the COVID-19 severity assessment13.
140 patients with COVID-19 participated in the retrospective research. The obtained clinical and laboratory data was supplemented by medical visualization data. A convolutional neural network carried out quantitative assessment of pulmonary foci on CT scans, which were confirmed by radiologists, and summarized the obtained scores13.
AI-CT rating system based on AI showed high forecasting accuracy with AUC of 0.929. This allowed for faster identification of patients in grave condition and conduction of oxygen therapy and lung ventilation13.
Artificial intelligence contributed greatly to combatting the pandemic, helping to solve problems beyond diagnosis only. The use of AI in COVID-19 treatment helped to repurpose medicines and accelerate vaccine development.
Researchers from the Radboud University Medical Center in the Netherlands considered the possibilities of a neural network to classify nodules when carrying out screening CT14.
A study published in Scientific Reports explores six types of nodules that can be differentiated on a CT scan by texture, form and intensity of the image. A physician has not only to establish the nature of a diagnostic finding, but also determine the possibility of its transfer to a malignant form. This can help to choose the correct tactics for further examinations14.
The researchers have developed a classification structure based on ConvNets convolutional neural networks. Each radiologist may assess a finding on a CT in different ways, so it was important for the researchers to apply standardized approached to neoplasm diagnosis. The system is based on Lung-RADS assessment categories and PanCan malignancy probability model14.
How was the research conducted14?
After training and checking the algorithm on the data of more than 1,000 patients, the researchers started testings.
The efficiency of the neural network was compared with the conclusions of one radiology technician and two radiologists with over 20 years of CT scan interpretation experience.
The diagnosticians looked through the scans and specified the nodule and its type.
The average accuracy of the neural network was comparable with the human accuracy: the results were 72.9% and 69.6% respectively. However, the researchers point out that both a physician and a computer can identify a finding on a CT mistakenly14.
CT Lungs service by SberMedAI company helps a radiologist to identify lung cancer at the early stage:
it analyzes tomography scans of chest organs;
then it detects minimal nodules in lung tissue, which size exceeds the limit of 4 millimeters;
it lights suspicious findings and draws attention of a diagnostician to it.
The algorithm can be used during a retrospective examination of patients with COVID-19. At the same time, the images, on which a physician has identified viral pneumonia, are analyzed again for neoplasms in lungs.
Google Health team has been engaged in the issue of lung cancer forecasting since 2017. Their machine learning model helped to make a step forward in early malignant tumor diagnosis. The results were published in 2019 in Nature Medicine15.
This is how the developed software can help physicians and patients15:
it can take into account the results of previous examinations;
it generates three-dimensional models;
it identifies minor nodules in lung tissue;
it allows for the assessment of general malignancy forecast.
The researchers provide promising figures15:
the computer detected 5% more cases of cancer than the team of six radiologists;
the number of false positive results, when healthy persons were mistakenly diagnosed with cancer, reduced by 11%;
AUC classification efficiency indicator achieved 94.4%.
The goal of the research was to increase the value of lung cancer screening by means of machine learning algorithms. The team plans to continue clinical researches worldwide15.
The use of artificial intelligence is a way to increase accuracy and speed of lung cancer detection. A physician can, combining their experience with the computing capabilities of cutting-edge technologies, answer the question «What to do next?» in favor of a patient.
C H McCollough and S Leng, Use of Artificial Intelligence in Computed Tomography Dose Optimisation // Ann ICRP. 2020 December;49:113-125. doi: 10.1177/0146645320940827.
CT Lungs ❘ SBERMEDAI: SberMedAI website. [Online]. URL: https://sbermed.ai/en/diagnostic-center/our-algorithms/ct-lungs/ (accessed 17/03/2022)
J.E. Soun, D.S. Chow, M. Nagamine, et al. Artificial Intelligence and Acute Stroke Imaging // AJNR Am J Neuroradiol. 2021 January;42(1):2-11. doi: 10.3174/ajnr.A6883.
Ilker Ozsahin, Boran Sekeroglu, Musa Sani Musa, et al. Review on Diagnosis of COVID-19 from Chest CT Images Using Artificial Intelligence // Comput Math Methods Med. 2020;2020:9756518. doi: 10.1155/2020/9756518.
Pedro Vinícius Staziaki, Di Wu, Jesse C Rayan, Irene Dixe de Oliveira Santo, et al. Machine Learning Combining CT Findings and Clinical Parameters Improves Prediction of Length of Stay and ICU Admission in Torso Trauma // Eur Radiol. 2021 July; 31 (7): 5434-5441. doi: 10.1007/s00330-020-07534-w.
The Impact of Artificial Intelligence on CT Imaging [Online]: Imagine Technology News. URL: https://www.itnonline.com/article/impact-artificial-intelligence-ct-imaging. (accessed 17/03/2022)
Iam Palatnik de Sousa, Marley M. B. R. Vellasco and Eduardo Costa da Silva, Explainable Artificial Intelligence for Bias Detection in COVID CT-Scan Classifiers // Sensors. 2021;21(16):5657. doi: 10.3390 / s21165657.
Samuel G Finlayson, Adarsh Subbaswamy, Karandeep Singh, et al. The Clinician and Dataset Shift in Artificial Intelligence // N Engl J Med. 2021 15 July;385(3):283-286. doi: 10.1056/NEJMc2104626.
Minjae Lee, Hyemi Kim and Hee-Joung Kim Sparse-View CT Reconstruction Based on Multi-Level Wavelet Convolution Neural Network // Phys Med. 2020 December; 80: 352-362. doi: 10.1016/j.ejmp.2020.11.021.
Maria Paola Belfiore, Fabrizio Urraro, Roberta Grassi et al. Artificial Intelligence to Codify Lung CT in Covid-19 Patients // Radiol Med. 2020 May;125(5):500-504. doi: 10.1007/s11547-020-01195-x.
Stephanie A. Harmon, Thomas H. Sanford, Sheng Xu et al. Artificial Intelligence for the Detection of COVID-19 Pneumonia on Chest CT Using Multinational Datasets // Nat Commun. 11, 4080 (2020). doi: 10.1038/s41467-020-17971-2.
Lin Li, Lixin Qin, Zeguo Xu et al. Using Artificial Intelligence to Detect COVID-19 and Community-acquired Pneumonia Based on Pulmonary CT: Evaluation of the Diagnostic Accuracy // Radiology. 2020;296(2):E65-E71. doi:10.1148/radiol.2020200905.
Yi Han, Su-Cheng Mu, Hai-Dong Zhang et al. Artificial Intelligence Computed Tomography Helps Evaluate the Severity of COVID-19 Patients: A Retrospective Study // World J Emerg Med. 2022;13(2):91-97. doi: 10.5847/wjem.j.1920-8642.2022.026.
Francesco Ciompi, Kaman Chung, Sarah J. van Riel, et al. Towards Automatic Pulmonary Nodule Management in Lung Cancer Screening with Deep Learning // Scientific Reports. 2017;7:46479. doi: 10.1038 / srep46479.
A Promising Step Forward for Predicting Lung Cancer [Online]: Google. URL: https://blog.google/technology/health/lung-cancer-prediction/. (accessed 28/01/2022)