5 minute read
Rama Asuri - Homework 6 AI in Health and Medicine
AI in health and medicine is achieving expert-level performance. In this paper, we examine two different examples where AI could detect and predict the malign tumors 1.
In the first example, we will cover Cervical cancer. Cervical cancer kills more women in India than in any other country. It is a preventable disease that kills 67000 women in India. Screening and detection can help reduce the number of deaths, but the challenge is the testing process, which takes enormous time. SRL diagnostics partnered with Microsoft to co-create an AI Network of Pathology to reduce cytopathologists and histopathologists' burden. Cytopathologists at SRL Diagnostics manually marked their observations. These observations were used as training data for Cervical Cancer Image Detection. However, there was a different challenge, the way cytopathologists examine different elements are unique even though they all have come to the same conclusion. This was because these experts may approach a problem from a different direction. The Manish Gupta, Principal Applied Researcher at Microsoft Azure Global Engineering, who worked closely with the team at SRL Diagnostics, said the idea was to create an AI algorithm that could identify areas that everybody was looking at and “create a consensus on the areas assessed.” Cytopathologists across multiple labs and locations annotated thousands of tile images of a cervical smear. They created discordant and concordant notes on each sample image. “The images for which annotations were found to be discordant — that is if they were viewed differently by three team members — were sent to senior cytopathologists for final analysis”. SRL Diagnostics has started an internal preview to use Cervical Cancer Image Detection API. The Cervical Cancer Image Detection API, which runs on Microsoft’s Azure, can quickly screen liquid-based cytology slide images to detect cervical cancer in the early stages and return insights to pathologists in labs. The AI model can now differentiate between normal and abnormal smear slides with accuracy and is currently under validation in labs. It can also classify smear slides based on the seven-subtypes of cervical cytopathological scale 1. Artificial intelligence can spot subtle patterns that can easily be missed by humans 2.
The second example is about detecting lung cancer. The survival rate is really high if lung cancer is detected during the early stages. Nevertheless, the problem is that it is difficult to do it manually when there are millions of 3D X-rays. Reviewing scans is done by a highly trained specialist, and a majority of the reviews result in no detection. Moreover, this is also monotonous work, which might lead to errors by the reviewers. The LUNA Grand Challenge is an open dataset with high-quality labels of patient CT scans. The gLUNA Grand Challenge encourages improvements in nodule detection by making it easy for teams to compete for high positions on the leader board. A project team can test the efficacy of their detection methods against standardized criteria 3.
Challenges and future of AI in Medical
AI models can do complex nonlinear relationships, fault tolerance, parallel distributed processing, and learning. With its ability to self-learn, concurrent processing of quantitative and qualitative knowledge, and validate the output from several clinical studies from many different fields, AI is used in different clinical medicine. It takes full advantage of the different aspects of clinical diversity and speaks to the current lack of objectivity and completeness. The application of AI helps train fresh out of school physicians in clinical diagnosis and decision-making. An increasing number of research papers report the accurate diagnosis and prognosis performance of the Machine Learning algorithm. Deep Learning techniques are transforming how radiologists interpret imaging data. These results may increase sensitivity and assure less number of false positives than radiologists. However, they drive the risk of overfitting the training data, resulting in a degradation in certain settings. Machine Learning involves a tradeoff between precision and intelligibility. More accurate models, such as boosted trees, random forests, and neural nets, are usually not intelligible, whereas more intelligible models, such as logistic regression, naive-Bayes, and single decision trees, often provide significantly worse accuracy. Recent advancements in vivo imaging, computational modeling, and animal modeling have identified barriers in the tumor microenvironment that interrupt therapy and promote tumor progression. Other risk factors identified from blood counts, red cell distribution width were used in a Machine Learning-based approach to generate a clinical data-driven prediction model capable of predicting acute myeloid leukemia 6–12 months before diagnosis with high specificity (98.2%) but low sensitivity (25.7%). Therefore, the application of AI in clinical cancer is likely to increase; the following challenges should be met in order for it to remain viable. AI technology faces remarkable challenges that must be resolved to guarantee its cancer diagnosis and prognosis. For example, medical imaging data must be transformed before it is used. It is crucial to extract features from the imaging data and process them. Medical interpretation needs further research because the models are tested based on weights and predict the output. Interdisciplinary personnel training and collaboration gaps must be filled through academic coursework and orgnizational trainings. A shift to Machine Learning statistical tools is critical for anyone practicing medicine in the 21st century 4.