Aimon Rahman, a third-year doctoral student in the Vision & Image Understanding (VIU) Lab in the Electrical and Computer Engineering Department, is making significant contributions to the field of medical artificial intelligence (AI). In her Hopkins Engineering Applications and Research Tutorials (HEART) course titled "Introduction to Deep Learning for Medical Imaging," Rahman introduces students to the practical applications of computer vision in medical image analysis.
In an interview with The News-Letter, Rahman shared insights on her recent publications, as well as opinions regarding the opportunities and challenges associated with medical AI.
Deep learning is a method of creating AI relying on neural networks, with many independent modeling layers. Their ability to handle large and intricate datasets makes them well-suited for tasks like medical imaging analysis.
Rahman’s work centers on training deep learning models to detect anomalies such as tumors and lesions across various modalities, including ultrasound, CT and MRI scans.
One of Rahman’s recent publications trains diffusion models to replicate the ''collective insights'' of doctors. Her paper, titled "Ambiguous medical image segmentation using diffusion models," includes an algorithm that can detect tumors given a set of images.
What sets Rahman’s work apart is her approach of countering the deterministic nature of AI by using a probabilistic model to mirror the diagnostic results of multiple doctors as collective insights.
"Let's say you have a segmentation model that analyzes medical images and segments out tumors. Whenever you have this kind of model, it's always deterministic, meaning you input an image, and it will provide a segmentation specific to that image. However, in a hospital setting, multiple radiologists may each have their unique diagnosis and segmentation for the same image,” she explained. “This is what I worked on last summer, where I developed a probabilistic model. When you input an image, the network will produce a variety of segmentation maps, mimicking the diagnostic results of different doctors."
Rahman also shed light on the potential applications of her work in health care. She emphasized the timely importance of AI and the abundance of resources available at Hopkins.
“Right now is a favorable time to delve into AI, and Hopkins offers extensive resources,“ she said. “My goal is to leverage computer vision to make a meaningful impact in health care. For instance, incorporating AI into health care can significantly reduce costs, which can be particularly beneficial for countries with limited resources and in point-of-care scenarios.”
Rahman underscored the technology's relevance in situations like wartime, refugee crises or natural disasters, where makeshift hospitals may lack health care professionals. Computer vision and deep learning can provide rapid and accurate diagnosis in such critical circumstances.
Looking toward the future, Rahman hopes that medical AI will make health care better and more affordable for all.
"Medical AI will be integrated in point-of-care architecture so that patients do not have to wait a long time in hospitals. The overarching aim is to use this technology to cut the health care cost and make standard health care more affordable in countries with few resources," she said.
However, Rahman acknowledges the challenges she faces, primarily concerning the scarcity of training data and the generalizability of the models, an issue compounded by the unavailability of private datasets.
"The biggest challenge in the medical domain is the limited dataset availability, as medical data are not publicly accessible,“ she said. “Obtaining data involves navigating ethical concerns and obtaining permissions. Moreover, data collection by medical professionals is time-consuming.”
One possible solution to these issues is the use of synthetic data — which itself is created by a neural network. However, when asked about the possibility of using synthetic datasets for medical training, Rahman expressed caution.
“To use a synthetic dataset, we need to have a bunch of data to begin with and train the model first, which is hard due to the aforementioned reasons. Furthermore, relying on healthy patient data to mimic severe cases can introduce bias into the dataset, as AI can only generate based on what it has seen before,'' she said. "We don’t want to give AI too much creative freedom so that it would create something that is impossible to occur in real life."