Written by: Abhay Mahajan
Edited by: Ryan Lee
Edited by: Ryan Lee
It is becoming increasingly evident that artificial intelligence (AI) will be at the center stage for many of our technological advancements in the future. With the rapid development of AI seen through novel AI art software, chatbots such as ChatGPT, and Google’s Bard, it is no surprise that artificial intelligence has eventually found its way into the field of medicine. Although the uses for AI in the medical field are vast and are continuing to grow, it is important to not simply focus on how AI can replace jobs, but rather how it can supplement work already being done.
Magnetic resonance imaging (MRI) scans are noninvasive and used to take a detailed look at the organs, bones, muscles, and blood vessels in the body, with much more clarity being shown for the images of soft tissue (John Hopkins 2023). Computerized Axial Tomography (CT), on the other hand yields many of the same benefits as MRI except that there is much less noise in the imaging, and CT has been shown to be better at detecting lung nodules in comparison to MRI (Kuriyama et al. 1990). This makes CT, especially with the combination of positron emission tomography (PET), a great option for detecting lesions within the lungs. Despite CT being one of the best options for the detection of lung nodules, a study by the Journal of Thoracic Imaging demonstrated that an average of 25% of lung nodules were missed by CT scans simply due to being outside of the gaze volume (Rubin 2016). This gaze volume is the region radiologists tend to view most based on their search pattern (Rubin 2016). Lesions lying outside of this gaze are thus more likely to be missed by the radiologists as they are not investigated as thoroughly. Even within this gaze volume, the sensitivity of lung nodule detection ranged from 47% to 84% (Rubin 2016). This wide range of sensitivity values underlines the potential purpose AI could serve in reducing the element of human error introduced in the reading of these CT scans.
Jordan Chamberlin and colleagues combined CT scans with two convolutional neural networks (CNN), a form of AI which learns about the common features between the samples presented in a large data set (Chamberlin et al. 2021; Intel 2021). In this study, one CNN was dedicated to coronary artery calcium volume analysis while the other was assigned to low dosage CT scans (LDCT) (Chamberlin et al. 2021). For the sake of this article, the focus will only be placed on the low dosage CT scans used to evaluate the lungs. Data was obtained from 117 participants (>18 years old). The method utilized nodule candidate generation, a software used to label certain lesions as potential cancerous nodules while also providing the respective probability scores (Chamberlin et al. 2021). Afterwards, these potential nodules were put through a false-positive reduction algorithm which aided in determining the likelihood of the candidates being true positives or not (Chamberlin et al. 2021). Combining both steps resulted in a score used to make the final decision of whether the lung lesions were nodules (Chamberlin et al. 2021).
To aid in determining if lung nodules were present, the lungs were segmented by lobes, with five lung lobes having their 3D CT volumes analyzed in order to determine the likelihood of a nodule being present in each lobe (Chamberlin et al. 2021). Determining the validity of the AI’s assessment was done through the use of two radiologists’ consensus based on the CT scan, which is the gold standard for determining the presence of a lung nodule (Chamberlin et al. 2021). If both the AI prediction and radiologists agreed on the presence of the nodule, the patient was considered a true positive, and in a case where both agreed that a nodule was not present, the patient was labeled a true negative (Chamberlin et al. 2021).
Consideration of variables such as sex, race, current smoking status, diabetes diagnosis, etc. had been conducted to determine prognostic value of the AI model (Chamberlin et al. 2021). When determining the strength of the aforementioned variables in determining one’s prognosis, variable importance in projection, or the impact of the variables used in a partial least squares was greater than 1 with a 95% confidence interval (Chamberlin et al. 2021). A VIP value greater than 1 means that the variable has a significant impact on the model (You 2015). Additionally, the 95% confidence interval means that there is only a 5% chance that the given range of data would exclude the individual strength of a predictor. The researchers had also found that the predictive value of variables such as demographics, clinical attributes, and risk factors when used with the AI model had an McFadden’s R-squared value of 0.142 for the prediction of pulmonary hospitalization and a value of 0.139 for predicting lung cancer (Chamberlin et al. 2021). These variables also aided in determining the likelihood of the nodules detected simply being false-positives (Chamberlin et al. 2021). It was determined that, in this case, the AI model was only significant when it comes to diagnosing lung cancer at a 1-year follow-up (Chamberlin et al. 2021). An issue, however, that seems to appear with the model is that it has a high sensitivity, but is low in terms of specificity, with 96% of the nodules detected in the lung LDCT’s having been found to be benign (Chamberlin et al. 2021). There is always a risk-gain assessment to be made in medicine, so it is very important to find ways in which we can accurately suspect cancer in the lungs without having to necessarily do an invasive procedure every time a nodule is spotted. Another limitation of the AI mentioned by the researchers themselves was that the lung nodules were significant in their lung cancer model but the presence of nodules does not exactly have a clear predictive contribution to the model (Chamberlin et al. 2021).
Artificial intelligence is expected to improve in terms of its reliability in the medical field. Already, a newer study in 2022 by Yaping Zhang et al. found that AI even assisted radiologists in detecting nodules in 247 of 250 patients when radiologists alone only detected nodules in 131 (Zhang et al. 2022). The future is uncertain yet hopeful for AI, as we strive to find ways in which our applications of machine learning algorithms can better improve the existing fields of today.
References
Convolutional Neural Networks (CNNs), Deep Learning, and Computer Vision [Internet]. Intel. [updated 2021, cited March 4, 2023]. Available from: https://www.intel.com/content/www/us/en/internet-of-things/computer-vision/convolutional-neural-networks.html
Fengqi You. 2015. Sustainability of Products, Processes, and Supply Chains [Internet]. 1st Edition. Elsevier. [cited March 4, 2023]. Available from: https://www.elsevier.com/books/T/A/9780444634726
Geoffrey D Rubin. 2016. Lung Nodule and Cancer Detection in CT Screening. Journal of Thoracic Imaging [Internet]. [cited March 4, 2023]. 30(2). Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4654704/
Jordan Chamberlin, Madison R. Kocher, Jeffrey Waltz, Madalyn Snoddy, Natalie F. C. Stringer, Joseph Stephenson, Pooyan Sahbaee, Puneet Sharma, Saikiran Rapaka, U. Joseph Schoepf, Andres F. Abadia, Jonathan Sperl, Phillip Hoelzer, Megan Mercer, Nayana Somayaji, Gilberto Aquino, Jeremy R. Burt 2021. Automated detection of lung nodules and coronary artery calcium using artificial intelligence on low-dose CT scans for lung cancer screening: accuracy and prognostic value. BMC Medicine [Internet]. [cited March 4, 2023]. 19(55). Available from: https://doi.org/10.1186/s12916-021-01928-3
K Kuriyama, T Kadota, C Kuroda. 1990. [CT and MR imaging in the evaluation and staging of lung cancer]. Gan To Kagaku Ryoho. [cited March 4, 2023]. Available from: https://pubmed.ncbi.nlm.nih.gov/2173493/
Magnetic Resonance Imaging (MRI) [Internet]. John Hopkins: [updated 2023, cited March 4, 2023]. Available from:https://www.hopkinsmedicine.org/health/treatment-tests-and-therapies/magnetic-resonance-imaging-mri#:~:text=MRI%20may%20be%20used%20instead,radiation%20during%20an%20MRI%20procedure
Yaping Zhang, Beibei Jiang, Lu Zhang, Marcel J W Greuter, Geertruida H de Bock Hao Zhang , Xueqian Xie. 2022. Lung Nodule Detectability of Artificial Intelligence-assisted CT Image Reading in Lung Cancer Screening. Current Medical Imaging [Internet]. [cited March 4, 2023]. 18(3). Available from: https://pubmed.ncbi.nlm.nih.gov/34365951/
Magnetic resonance imaging (MRI) scans are noninvasive and used to take a detailed look at the organs, bones, muscles, and blood vessels in the body, with much more clarity being shown for the images of soft tissue (John Hopkins 2023). Computerized Axial Tomography (CT), on the other hand yields many of the same benefits as MRI except that there is much less noise in the imaging, and CT has been shown to be better at detecting lung nodules in comparison to MRI (Kuriyama et al. 1990). This makes CT, especially with the combination of positron emission tomography (PET), a great option for detecting lesions within the lungs. Despite CT being one of the best options for the detection of lung nodules, a study by the Journal of Thoracic Imaging demonstrated that an average of 25% of lung nodules were missed by CT scans simply due to being outside of the gaze volume (Rubin 2016). This gaze volume is the region radiologists tend to view most based on their search pattern (Rubin 2016). Lesions lying outside of this gaze are thus more likely to be missed by the radiologists as they are not investigated as thoroughly. Even within this gaze volume, the sensitivity of lung nodule detection ranged from 47% to 84% (Rubin 2016). This wide range of sensitivity values underlines the potential purpose AI could serve in reducing the element of human error introduced in the reading of these CT scans.
Jordan Chamberlin and colleagues combined CT scans with two convolutional neural networks (CNN), a form of AI which learns about the common features between the samples presented in a large data set (Chamberlin et al. 2021; Intel 2021). In this study, one CNN was dedicated to coronary artery calcium volume analysis while the other was assigned to low dosage CT scans (LDCT) (Chamberlin et al. 2021). For the sake of this article, the focus will only be placed on the low dosage CT scans used to evaluate the lungs. Data was obtained from 117 participants (>18 years old). The method utilized nodule candidate generation, a software used to label certain lesions as potential cancerous nodules while also providing the respective probability scores (Chamberlin et al. 2021). Afterwards, these potential nodules were put through a false-positive reduction algorithm which aided in determining the likelihood of the candidates being true positives or not (Chamberlin et al. 2021). Combining both steps resulted in a score used to make the final decision of whether the lung lesions were nodules (Chamberlin et al. 2021).
To aid in determining if lung nodules were present, the lungs were segmented by lobes, with five lung lobes having their 3D CT volumes analyzed in order to determine the likelihood of a nodule being present in each lobe (Chamberlin et al. 2021). Determining the validity of the AI’s assessment was done through the use of two radiologists’ consensus based on the CT scan, which is the gold standard for determining the presence of a lung nodule (Chamberlin et al. 2021). If both the AI prediction and radiologists agreed on the presence of the nodule, the patient was considered a true positive, and in a case where both agreed that a nodule was not present, the patient was labeled a true negative (Chamberlin et al. 2021).
Consideration of variables such as sex, race, current smoking status, diabetes diagnosis, etc. had been conducted to determine prognostic value of the AI model (Chamberlin et al. 2021). When determining the strength of the aforementioned variables in determining one’s prognosis, variable importance in projection, or the impact of the variables used in a partial least squares was greater than 1 with a 95% confidence interval (Chamberlin et al. 2021). A VIP value greater than 1 means that the variable has a significant impact on the model (You 2015). Additionally, the 95% confidence interval means that there is only a 5% chance that the given range of data would exclude the individual strength of a predictor. The researchers had also found that the predictive value of variables such as demographics, clinical attributes, and risk factors when used with the AI model had an McFadden’s R-squared value of 0.142 for the prediction of pulmonary hospitalization and a value of 0.139 for predicting lung cancer (Chamberlin et al. 2021). These variables also aided in determining the likelihood of the nodules detected simply being false-positives (Chamberlin et al. 2021). It was determined that, in this case, the AI model was only significant when it comes to diagnosing lung cancer at a 1-year follow-up (Chamberlin et al. 2021). An issue, however, that seems to appear with the model is that it has a high sensitivity, but is low in terms of specificity, with 96% of the nodules detected in the lung LDCT’s having been found to be benign (Chamberlin et al. 2021). There is always a risk-gain assessment to be made in medicine, so it is very important to find ways in which we can accurately suspect cancer in the lungs without having to necessarily do an invasive procedure every time a nodule is spotted. Another limitation of the AI mentioned by the researchers themselves was that the lung nodules were significant in their lung cancer model but the presence of nodules does not exactly have a clear predictive contribution to the model (Chamberlin et al. 2021).
Artificial intelligence is expected to improve in terms of its reliability in the medical field. Already, a newer study in 2022 by Yaping Zhang et al. found that AI even assisted radiologists in detecting nodules in 247 of 250 patients when radiologists alone only detected nodules in 131 (Zhang et al. 2022). The future is uncertain yet hopeful for AI, as we strive to find ways in which our applications of machine learning algorithms can better improve the existing fields of today.
References
Convolutional Neural Networks (CNNs), Deep Learning, and Computer Vision [Internet]. Intel. [updated 2021, cited March 4, 2023]. Available from: https://www.intel.com/content/www/us/en/internet-of-things/computer-vision/convolutional-neural-networks.html
Fengqi You. 2015. Sustainability of Products, Processes, and Supply Chains [Internet]. 1st Edition. Elsevier. [cited March 4, 2023]. Available from: https://www.elsevier.com/books/T/A/9780444634726
Geoffrey D Rubin. 2016. Lung Nodule and Cancer Detection in CT Screening. Journal of Thoracic Imaging [Internet]. [cited March 4, 2023]. 30(2). Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4654704/
Jordan Chamberlin, Madison R. Kocher, Jeffrey Waltz, Madalyn Snoddy, Natalie F. C. Stringer, Joseph Stephenson, Pooyan Sahbaee, Puneet Sharma, Saikiran Rapaka, U. Joseph Schoepf, Andres F. Abadia, Jonathan Sperl, Phillip Hoelzer, Megan Mercer, Nayana Somayaji, Gilberto Aquino, Jeremy R. Burt 2021. Automated detection of lung nodules and coronary artery calcium using artificial intelligence on low-dose CT scans for lung cancer screening: accuracy and prognostic value. BMC Medicine [Internet]. [cited March 4, 2023]. 19(55). Available from: https://doi.org/10.1186/s12916-021-01928-3
K Kuriyama, T Kadota, C Kuroda. 1990. [CT and MR imaging in the evaluation and staging of lung cancer]. Gan To Kagaku Ryoho. [cited March 4, 2023]. Available from: https://pubmed.ncbi.nlm.nih.gov/2173493/
Magnetic Resonance Imaging (MRI) [Internet]. John Hopkins: [updated 2023, cited March 4, 2023]. Available from:https://www.hopkinsmedicine.org/health/treatment-tests-and-therapies/magnetic-resonance-imaging-mri#:~:text=MRI%20may%20be%20used%20instead,radiation%20during%20an%20MRI%20procedure
Yaping Zhang, Beibei Jiang, Lu Zhang, Marcel J W Greuter, Geertruida H de Bock Hao Zhang , Xueqian Xie. 2022. Lung Nodule Detectability of Artificial Intelligence-assisted CT Image Reading in Lung Cancer Screening. Current Medical Imaging [Internet]. [cited March 4, 2023]. 18(3). Available from: https://pubmed.ncbi.nlm.nih.gov/34365951/
Proudly powered by Weebly