Massive amounts of medical data—from imaging scans, patient visits, and electronic health records—provides an unprecedented opportunity for training health care AI models. But these data present their own challenges: Training AI models on continually arriving data is impractical at best and catastrophic at worst, as AI tends to forget old data when it learns new data.
A team from the Johns Hopkins University’s Computational Cognition, Vision, and Learning research group has developed a potential solution: an efficient online learning method that significantly reduces AI models’ tendency to forget older data by ensuring the models learn from only the most significant data samples. The team will present its work at the 27th International Conference on Medical Image Computing and Computer-Assisted Intervention, to be held October 6–10 in Marrakesh, Morocco.
The researchers’ method first uses text encoding to understand and categorize the medical data that an AI model is being trained on.
“This means that the AI can recognize and work with classes of pathologies or organs just by their descriptions,” explains Yu-Cheng Chou, a PhD student in the Department of Computer Science and one of the members on the team.
The group’s method then utilizes three different types of memory to ensure maximum learning efficiency. First, it uses linear memory to store recent samples as new data become available, enabling continuous training without having to revisit old data or make inefficient passes through the whole dataset.
It then uses dynamic memory to choose data samples based on uniqueness, enhancing the diversity of the training data. Preserving samples from different time periods and sources prevents the model from forgetting older data, solving one of the key challenges in training on massive datasets, says Chou.
Finally, their method makes use of selective memory to learn from the most challenging samples—those that the model is most uncertain about segmenting—to improve performance without requiring extra computational resources.
“Inspired by human learning patterns, we prioritize samples based on their significance, constructing a memory that emphasizes the most challenging samples while discarding the easier ones,” the researchers write.
In a series of tests on two different CT scan datasets, the team shows that its final learning method both surpasses current training paradigms in terms of performance and efficiency and significantly reduces catastrophic forgetting.
“Our findings provide a scalable, efficient way to train AI models on massive medical data streams of potentially infinite length—which is crucial for real-world applications, such as those in evolving clinical scenarios,” says Chou.
The team plans to further refine its learning method to handle more diverse and even larger-scale datasets by exploring fine-grained and annotation-free learning techniques.
Learn more about their project here.
Additional authors of this work include Zongwei Zhou, an assistant research scientist in the Department of Computer Science, and Alan Yuille, the Bloomberg Distinguished Professor of Computational Cognitive Science.