Document Type
Poster Session
Department
Computer Sciences
Faculty Mentor
Bruce MacLeod, PhD
Keywords
medical image segmentation, few-shot learning, deep learning
Abstract
Deep learning models can be difficult to train because they require large amounts of data, which we usually do not have or are too expensive to get or annotate. To overcome this problem, we can use few-shot meta-learning, which allows us to train deep learning models with little data. Using a few examples, meta-learning, or learning-to-learn, aims to use the experience learned during training to generalize to unknown tasks. Medical imaging is an industry where it is particularly useful, as there is limited publicly available data due to patient privacy concerns and annotating costs.
This project examines how meta-learning performs on medical imaging tasks by using the Medical Segmentation Decathlon dataset, which consists of ten different segmentation tasks for Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) images. This dataset is separated into a development set, consisting of seven tasks, and a mystery set consisting of three tasks to be used later. Using the U-Net architecture, we intend to create a pipeline for deep meta-learning that will train a model on the development set that generalizes effectively enough to handle unseen tasks. Doing so would provide a model that is flexible enough to adapt to unseen tasks without needing to be retrained. To simplify the process, at first, we focus on only 2D MRI and CT slices instead of 3D and utilize these slices to train models based on the Model-Agnostic Meta Learning (MAML) and Reptile meta-learning algorithms. Currently, the best model to be achieved using MAML is a 3-way 7-shot U-Net model with an intersection over union (IoU) of 0.241, while the best Reptile model is a 3-way 10-shot U-Net model with an IoU of 0.49.
Utilizing Few-Shot Meta Learning Algorithms for Medical Image Segmentation
Deep learning models can be difficult to train because they require large amounts of data, which we usually do not have or are too expensive to get or annotate. To overcome this problem, we can use few-shot meta-learning, which allows us to train deep learning models with little data. Using a few examples, meta-learning, or learning-to-learn, aims to use the experience learned during training to generalize to unknown tasks. Medical imaging is an industry where it is particularly useful, as there is limited publicly available data due to patient privacy concerns and annotating costs.
This project examines how meta-learning performs on medical imaging tasks by using the Medical Segmentation Decathlon dataset, which consists of ten different segmentation tasks for Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) images. This dataset is separated into a development set, consisting of seven tasks, and a mystery set consisting of three tasks to be used later. Using the U-Net architecture, we intend to create a pipeline for deep meta-learning that will train a model on the development set that generalizes effectively enough to handle unseen tasks. Doing so would provide a model that is flexible enough to adapt to unseen tasks without needing to be retrained. To simplify the process, at first, we focus on only 2D MRI and CT slices instead of 3D and utilize these slices to train models based on the Model-Agnostic Meta Learning (MAML) and Reptile meta-learning algorithms. Currently, the best model to be achieved using MAML is a 3-way 7-shot U-Net model with an intersection over union (IoU) of 0.241, while the best Reptile model is a 3-way 10-shot U-Net model with an IoU of 0.49.
