In the 21st century, there is no doubt that technology has played an increasingly integral role in facilitating daily life. Within the field of medicine, integrating technology with healthcare practices has become especially vital for patient care. For example, during the COVID-19 pandemic, technology was utilized in the form of telemedicine visits to deliver care to immunocompromised patient populations, and infrared scanners were used in public areas to identify symptomatic individuals and prevent potential COVID-19 outbreaks (Budd et. al, 2020). Recently, one of the most intriguing areas of healthcare technological innovation has been the use of artificial intelligence (AI) to identify salient features in radiological images of the brain to improve early diagnosis and treatment in patients with life-threatening conditions. Though these AI algorithms are still being developed and tested, they could potentially have a significant impact on improving patient outcomes if scaled to widespread clinical implementation. This article will discuss the technological framework for such AI, explore two specific applications of AI which directly relate to my current research projects, and discuss the ethical implications of such technology.
Neural Networks
Radiological scans, such as computed tomography (CT) and magnetic resonance imaging (MRI) scans, are primarily analyzed with machine learning neural networks that process images using computer vision. These models are trained on previously labeled images in order to make predictions about structure classification and segmentation as well as help render 3D models for the radiologist (Sajid, 2023). The architecture of neural networks is fascinating since—as the name suggests—it is inspired by the networks of the human brain. Just like the brain is composed of interconnected neurons, neural networks consist of multiple layers of digital neurons. Specifically, these digital neurons are functions that transform a given input into an output by assigning the former a particular weight. Collectively, a layer of neurons acts as a filter that modifies the previous neuron layer’s output into related but slightly different information. Within radiological scans, the goal of neural networks is to obtain the right set of filters to funnel the input information from the image scan into a desired output, like the classification of a specific area on the image as a brain ventricle. This is achieved either by altering the number of layers and neurons or by modifying the weights on the inputs and interconnectivity of the neurons themselves. For each training image, the neural network makes a random output guess with its current weights and connections; it then uses the image label—the “answer key”—to evaluate the discrepancy or loss from the desired result and adjusts the neuron properties appropriately. Eventually, through this iterative process over the training image set, the neural network optimizes its neurons to achieve reasonable accuracy and can become a powerful aid in practice (IBM, 2023).
Medical Applications of Neural Networks
There are a myriad of neural network applications for different types of medical imaging, but I will discuss two specific examples involving brain CT scans from my current research. There are a number of diseases that can occur in the brain, and one particularly dangerous condition is intracerebral hemorrhage (ICH), or bleeding in the brain tissue due to a ruptured blood vessel. A spontaneous increase in the blood may damage brain tissue by leading to hematoma expansion and secondary injury including perihematomal edema; these effects correlated with worse functional outcomes and increased mortality risk (Kellner et. al 2021). As such, early surgical evacuation of ICH can be critical in improving patient outcomes. For small hemorrhages, however, the risks of invasive surgical procedures may outweigh the potential benefits. Therefore, timely and accurate estimation of ICH volume from the initial CT scan is paramount. Typically, using the “ABC/2” formula on CT imaging is the clinical gold standard, but this method can be inaccurate and takes time due to reviewer assessment. On the other hand, semi-autonomous segmentation (SAS) is very accurate in hemorrhage volume estimation but is not used clinically since the slice-by-slice analysis requires even more time. Thus, having a neural network model which can quickly segment a brain CT scan of a patient with hemorrhage and accurately compute hematoma volume would be beneficial in determining whether or not surgery is necessary. In this study, we looked at the Viz.AI segmentation algorithm and found it to be more accurate in estimating hematoma volume than the ABC/2 method and similar in accuracy to the SAS method; there was also a significant amount of time saved by automating segmentation through the network as compared to manually segmenting each slice of the scan (Odland, 2023). These results justify the use of Viz.AI in ICH triage workflow and are a promising indicator of AI’s efficacy in the healthcare setting.
Another application of neural networks in imaging is in the prediction of neurological outcomes after cardiac arrest. 80% of resuscitated cardiac arrest patients do not regain immediate consciousness after circulation restoration, and several may remain comatose for up to several weeks. Since nearly two-thirds of comatose patients die from hypoxic-ischemic brain injury, it is important to determine the likelihood of neurological recovery early on to determine the course of care (Sandroni et. al, 2018). An important prognostic indicator has typically been the Glasgow Coma Scale; however, a lower gray-to-white matter ratio has also been demonstrated to be a good indicator of severe cerebral edema and mortality (Wu, 2011). Currently, I am working to develop my own neural network model which uses brain CT scans to determine the gray-to-white matter ratio (as determined by mean Hounsfield Units of the caudate nucleus head divided by mean Hounsfield Units of the posterior limb of the internal capsule) in order to predict the likelihood of recovery for currently comatose cardiac arrest patients. Such automation of outcome prediction may help better inform clinicians, patients, and patient families of the optimal direction of treatment earlier on in the process, which could have emotional benefits for the patient and their family and economic benefits for the hospital.
Given the power of neural networks in the above applications, it may seem that such algorithms are a panacea for early disease diagnosis and treatment. However, there are important ethical questions that must be addressed before we can fully embrace this technology. It is important to consider that though a model may be trained to a particular accuracy, it is hard to explain how or why it arrived at its results. Oftentimes, these networks involve complex mathematical operations for over thousands of neurons that cannot be easily understood. As the interpretability of results is crucial for clinical integration and the slightest mistake can have immense consequences, such as the misdiagnosis of cancer, some form of logic and reasoning must be established before algorithms can be put into clinical practice. Another potential drawback is that these models require vast medical datasets of personal records to train these models--some point out that this is a breach of privacy (Sajid, 2023). But despite these moral and ethical apprehensions, it is undeniable that the level of innovation in healthcare treatment with AI is both unprecedented and exciting.
About the Author
Dylan Wu is a rising sophomore at Harvard College, concentrating in Applied Math, pursuing a joint Masters in Music with the New England Conservatory.
References
Neural Networks
Radiological scans, such as computed tomography (CT) and magnetic resonance imaging (MRI) scans, are primarily analyzed with machine learning neural networks that process images using computer vision. These models are trained on previously labeled images in order to make predictions about structure classification and segmentation as well as help render 3D models for the radiologist (Sajid, 2023). The architecture of neural networks is fascinating since—as the name suggests—it is inspired by the networks of the human brain. Just like the brain is composed of interconnected neurons, neural networks consist of multiple layers of digital neurons. Specifically, these digital neurons are functions that transform a given input into an output by assigning the former a particular weight. Collectively, a layer of neurons acts as a filter that modifies the previous neuron layer’s output into related but slightly different information. Within radiological scans, the goal of neural networks is to obtain the right set of filters to funnel the input information from the image scan into a desired output, like the classification of a specific area on the image as a brain ventricle. This is achieved either by altering the number of layers and neurons or by modifying the weights on the inputs and interconnectivity of the neurons themselves. For each training image, the neural network makes a random output guess with its current weights and connections; it then uses the image label—the “answer key”—to evaluate the discrepancy or loss from the desired result and adjusts the neuron properties appropriately. Eventually, through this iterative process over the training image set, the neural network optimizes its neurons to achieve reasonable accuracy and can become a powerful aid in practice (IBM, 2023).
Medical Applications of Neural Networks
There are a myriad of neural network applications for different types of medical imaging, but I will discuss two specific examples involving brain CT scans from my current research. There are a number of diseases that can occur in the brain, and one particularly dangerous condition is intracerebral hemorrhage (ICH), or bleeding in the brain tissue due to a ruptured blood vessel. A spontaneous increase in the blood may damage brain tissue by leading to hematoma expansion and secondary injury including perihematomal edema; these effects correlated with worse functional outcomes and increased mortality risk (Kellner et. al 2021). As such, early surgical evacuation of ICH can be critical in improving patient outcomes. For small hemorrhages, however, the risks of invasive surgical procedures may outweigh the potential benefits. Therefore, timely and accurate estimation of ICH volume from the initial CT scan is paramount. Typically, using the “ABC/2” formula on CT imaging is the clinical gold standard, but this method can be inaccurate and takes time due to reviewer assessment. On the other hand, semi-autonomous segmentation (SAS) is very accurate in hemorrhage volume estimation but is not used clinically since the slice-by-slice analysis requires even more time. Thus, having a neural network model which can quickly segment a brain CT scan of a patient with hemorrhage and accurately compute hematoma volume would be beneficial in determining whether or not surgery is necessary. In this study, we looked at the Viz.AI segmentation algorithm and found it to be more accurate in estimating hematoma volume than the ABC/2 method and similar in accuracy to the SAS method; there was also a significant amount of time saved by automating segmentation through the network as compared to manually segmenting each slice of the scan (Odland, 2023). These results justify the use of Viz.AI in ICH triage workflow and are a promising indicator of AI’s efficacy in the healthcare setting.
Another application of neural networks in imaging is in the prediction of neurological outcomes after cardiac arrest. 80% of resuscitated cardiac arrest patients do not regain immediate consciousness after circulation restoration, and several may remain comatose for up to several weeks. Since nearly two-thirds of comatose patients die from hypoxic-ischemic brain injury, it is important to determine the likelihood of neurological recovery early on to determine the course of care (Sandroni et. al, 2018). An important prognostic indicator has typically been the Glasgow Coma Scale; however, a lower gray-to-white matter ratio has also been demonstrated to be a good indicator of severe cerebral edema and mortality (Wu, 2011). Currently, I am working to develop my own neural network model which uses brain CT scans to determine the gray-to-white matter ratio (as determined by mean Hounsfield Units of the caudate nucleus head divided by mean Hounsfield Units of the posterior limb of the internal capsule) in order to predict the likelihood of recovery for currently comatose cardiac arrest patients. Such automation of outcome prediction may help better inform clinicians, patients, and patient families of the optimal direction of treatment earlier on in the process, which could have emotional benefits for the patient and their family and economic benefits for the hospital.
Given the power of neural networks in the above applications, it may seem that such algorithms are a panacea for early disease diagnosis and treatment. However, there are important ethical questions that must be addressed before we can fully embrace this technology. It is important to consider that though a model may be trained to a particular accuracy, it is hard to explain how or why it arrived at its results. Oftentimes, these networks involve complex mathematical operations for over thousands of neurons that cannot be easily understood. As the interpretability of results is crucial for clinical integration and the slightest mistake can have immense consequences, such as the misdiagnosis of cancer, some form of logic and reasoning must be established before algorithms can be put into clinical practice. Another potential drawback is that these models require vast medical datasets of personal records to train these models--some point out that this is a breach of privacy (Sajid, 2023). But despite these moral and ethical apprehensions, it is undeniable that the level of innovation in healthcare treatment with AI is both unprecedented and exciting.
About the Author
Dylan Wu is a rising sophomore at Harvard College, concentrating in Applied Math, pursuing a joint Masters in Music with the New England Conservatory.
References
- Budd, J., Miller, B. S., Manning, E. M., Lampos, V., Zhuang, M., Edelstein, M., ... & McKendry, R. A. (2020). Digital technologies in the public-health response to COVID-19. Nature medicine, 26(8), 1183-1192.
- IBM (2023). What Are Neural Networks? Ibm.com.
- Kellner, C. P., Schupper, A. J., & Mocco, J. (2021). Surgical Evacuation of Intracerebral Hemorrhage: The Potential Importance of Timing. Stroke, 52(10), 3391–3398.
- Odland, I. C. (2023, 02-08). Artificial Intelligence-driven Automated Intracerebral Hemorrhage Volume Calculation Is More Accurate Than ABC/2. International Stroke Confidence, Dallas TX.
- Sajid, H. (2023). Ai in Radiology: Pros & Cons, Applications, and 4 Examples. V7.com.
- Sandroni, C., D’Arrigo, S. & Nolan, J.P. (2018). Prognostication after cardiac arrest. Crit Care 22, 150.
- Wu, O., Batista, L. M., Lima, F. O., Vangel, M. G., Furie, K. L., & Greer, D. M. (2011). Predicting clinical outcome in comatose cardiac arrest patients using early noncontrast computed tomography. Stroke, 42(4), 985–992.
