Could we provide AI with basic medical training? Could we take them through medical terminology classes and end up with models that can ‘reason’ just like the doctor? Incorporating Knowledge-enhanced Bottlenecks into AI programmes designed for radiology could help them do just that.
Unlike human doctors, algorithms don’t suffer from burnout and can work on large volumes of work in a short time.
AI is already here, how do we make it a tool rather than a replacement?
Artificial intelligence has taken root in medicine, especially in the field of radiology. A new algorithm developed by researchers at the University of Pennsylvania aims to correct the shortcomings of existing AI models used in interpreting radiological scans. The algorithm aims to provide more accurate diagnoses by referring to various sources of knowledge such as research papers and medical textbooks. The developers aim to equip the AI model with more comprehensive knowledge in the field of medicine
Currently, the available AI models for image interpretation are mostly based on machine learning techniques that don’t quite reveal the process of interpretation. Generally, an AI system is exposed to a large data set. It uses the data to find repeated patterns that it can use to come up with a diagnosis.
For example, when the AI is shown lots of X-rays with features of pneumonia, it learns the patterns and retains them. This approach has been found to fall short in some circumstances. For example, when the AI is subjected to data from a different group of people, it is practically thrown off balance. An AI model trained on interpreting images of male patients might have trouble doing the same for female patients because of the anatomical differences between men and women.
Can we train AI like we train doctors?
Yue Yang and his colleagues at the University of Pennsylvania have made an artificial intelligence system that worksreliably in different medical settings. They posit that when the AI model is trained just like a human doctor, it will be more accurate in detecting pathologies acrossdifferent scenarios.
A radiologist can be transferred from Texas to Toronto and work perfectly fine. Save for some cultural adjustments and maybe a change in weather patterns, we expect them to find few challenges adjusting to the job. Their training and previous work will help them settle in and continue working. Ideally this should also be the case with a well-trained AI model when it is exposed to varying demographics. To achieve this, the scientists have curated a method of training the AI more thoughtfully.
Could adding Knowledge-enhanced Bottlenecks remove learning blockages for AI in radiology?
Rather than just feeding the program with data and recieving an output from the patterns seen in the large volumes of input, the AI should be trained to recognise and distinguish groupings within the data A human doctor has to undergo years of training before they are board-certified radiologists. When the radiologist pens a report of a scan, it is the product of human anatomy classes, countless sessions on dissection, their experiences in exams and the many clinical scenarios that they have come across through the journey.
Introducing KnoBo
The new model known as Knowledge-enhanced Bottlenecks (KnoBo) builds up the AI with the necessary medical knowledge. The researchers utilized a pool of resources including PubMed and Statpearls which have a host of relevant medical information. This way, the conclusion of the AI model is more rationalized. For instance, it was able to identify an X-ray of an infected lung based on the characteristic feature visualized ̶ the presence of ground-glass opacities. Knowledge-enhanced Bottlenecks could be the next revolution in AI for radiology.
Speaking to the news segment of Penn Engineering, Yue Yang, a doctoral student and first author of the paper states, “when reading an X-ray, medical students and doctors ask, is the lung clear, is the heart a normal size? The model will rely on similar factors to the ones humans use when making a decision.”
KnoBo utilizes the concept bottleneck model which is a way of creating more interpretable neural networks for AI. To achieve these high levels of interpretability, the model incorporates prior knowledge in medicine which then forms a basis of its image interpretation. Yue Yang and co also found that PubMed as a source of prior medical knowledge had great performance in terms of providing data with varying attributes and it enhanced the overall performance of the AI model.
We Aren’t Doing Away with Radiologists Just Yet
While the shortage of radiologists continues to bite, AI can’t be a foolproof solution in the foreseeable future. Even with advanced models like KnoBo, limitations still prevail. For example, the model as described by Yue Yang and his team still falls short when it comes to rare conditions. What’s more, novel medical conditions with unique radiological features can still arise, just like COVID-19, requiring the AI to relearn new patterns and come up with solutions.
We will still need a human doctor to look at the scans and come up with a final diagnosis. In the meantime, it is safe to say that AI will have to work with humans until we hit high levels of perfection.
The concept of equipping AI models with prior medical knowledge seems to be a promising venture for radiology and medicine at large. It is assuring to know that AI models will now have broader and a sturdier knowledge base unlike previously when they had to ‘cut corners’ to finding a solution. With some tweaks and advancements, concepts like KnoBo can be utilized in a wider scope including other areas of medicine and science at large to provide more robust solutions.
References
Scheffler, I. (2024) Training medical AI with knowledge, not shortcuts, Penn Engineering at University of Pennsylvania. Available at: https://ai.seas.upenn.edu/news/training-medical-ai-with-knowledge-not-shortcuts/ (Accessed: 01 December 2024).
Yang, Y., Gandhi, M., Wang, Y., Wu, Y., Yao, M.S., Callison-Burch, C., Gee, J.C. and Yatskar, M., 2024. A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis. arXiv preprint arXiv:2405.14839.
Could we provide AI with basic medical training? Could we take them through medical terminology classes and end up with models that can ‘reason’ just like the doctor? Incorporating Knowledge-enhanced Bottlenecks into AI programmes designed for radiology could help them do just that. Unlike human doctors, algorithms don’t suffer from burnout and can work on large volumes of work in a short time. Artificial intelligence has taken root in medicine, especially in the field of radiology. A new algorithm developed by researchers at the University of Pennsylvania aims to correct the shortcomings of existing AI models used in interpreting radiological scans. The algorithm aims to provide more accurate diagnoses by referring to various sources of knowledge such as research papers and medical textbooks. The developers aim to equip the AI model with more comprehensive knowledge in the field of medicine Currently, the available AI models for image interpretation are mostly based on machine learning techniques that don’t quite reveal the process of interpretation. Generally, an AI system is exposed to a large data set. It uses the data to find repeated patterns that it can use to come up with a diagnosis. For example, when the AI is shown lots of X-rays with features of pneumonia, it learns the patterns and retains them. This approach has been found to fall short in some circumstances. For example, when the AI is subjected to data from a different group of people, it is practically thrown off balance. An AI model trained on interpreting images of male patients might have trouble doing the same for female patients because of the anatomical differences between men and women. Yue Yang and his colleagues at the University of Pennsylvania have made an artificial intelligence system that worksreliably in different medical settings. They posit that when the AI model is trained just like a human doctor, it will be more accurate in detecting pathologies acrossdifferent scenarios. A radiologist can be transferred from Texas to Toronto and work perfectly fine. Save for some cultural adjustments and maybe a change in weather patterns, we expect them to find few challenges adjusting to the job. Their training and previous work will help them settle in and continue working. Ideally this should also be the case with a well-trained AI model when it is exposed to varying demographics. To achieve this, the scientists have curated a method of training the AI more thoughtfully. Rather than just feeding the program with data and recieving an output from the patterns seen in the large volumes of input, the AI should be trained to recognise and distinguish groupings within the data A human doctor has to undergo years of training before they are board-certified radiologists. When the radiologist pens a report of a scan, it is the product of human anatomy classes, countless sessions on dissection, their experiences in exams and the many clinical scenarios that they have come across through the journey. The new model known as Knowledge-enhanced Bottlenecks (KnoBo) builds up the AI with the necessary medical knowledge. The researchers utilized a pool of resources including PubMed and Statpearls which have a host of relevant medical information. This way, the conclusion of the AI model is more rationalized. For instance, it was able to identify an X-ray of an infected lung based on the characteristic feature visualized ̶ the presence of ground-glass opacities. Knowledge-enhanced Bottlenecks could be the next revolution in AI for radiology. Speaking to the news segment of Penn Engineering, Yue Yang, a doctoral student and first author of the paper states, “when reading an X-ray, medical students and doctors ask, is the lung clear, is the heart a normal size? The model will rely on similar factors to the ones humans use when making a decision.” KnoBo utilizes the concept bottleneck model which is a way of creating more interpretable neural networks for AI. To achieve these high levels of interpretability, the model incorporates prior knowledge in medicine which then forms a basis of its image interpretation. Yue Yang and co also found that PubMed as a source of prior medical knowledge had great performance in terms of providing data with varying attributes and it enhanced the overall performance of the AI model. While the shortage of radiologists continues to bite, AI can’t be a foolproof solution in the foreseeable future. Even with advanced models like KnoBo, limitations still prevail. For example, the model as described by Yue Yang and his team still falls short when it comes to rare conditions. What’s more, novel medical conditions with unique radiological features can still arise, just like COVID-19, requiring the AI to relearn new patterns and come up with solutions. We will still need a human doctor to look at the scans and come up with a final diagnosis. In the meantime, it is safe to say that AI will have to work with humans until we hit high levels of perfection. The concept of equipping AI models with prior medical knowledge seems to be a promising venture for radiology and medicine at large. It is assuring to know that AI models will now have broader and a sturdier knowledge base unlike previously when they had to ‘cut corners’ to finding a solution. With some tweaks and advancements, concepts like KnoBo can be utilized in a wider scope including other areas of medicine and science at large to provide more robust solutions. Scheffler, I. (2024) Training medical AI with knowledge, not shortcuts, Penn Engineering at University of Pennsylvania. Available at: https://ai.seas.upenn.edu/news/training-medical-ai-with-knowledge-not-shortcuts/ (Accessed: 01 December 2024). Yang, Y., Gandhi, M., Wang, Y., Wu, Y., Yao, M.S., Callison-Burch, C., Gee, J.C. and Yatskar, M., 2024. A Textbook Remedy for Domain Shifts: Knowledge Priors for Medical Image Analysis. arXiv preprint arXiv:2405.14839.
AI is already here, how do we make it a tool rather than a replacement?
Can we train AI like we train doctors?
Could adding Knowledge-enhanced Bottlenecks remove learning blockages for AI in radiology?
Introducing KnoBo
We Aren’t Doing Away with Radiologists Just Yet
References
