Multiscale deep reinforcement machine learning for n-dimensional segmentation in medical imaging

Abstract

Multiscale deep reinforcement machine learning for N-dimensional segmentation in medical imaging. Multiscale deep reinforcement learning produces multiscale deep reinforcement models for multidimensional (e.g., 3D) segmentation of objects (22). In this context, segmentation is formulated as learning an image-driven policy (38) for shape evolution (40) that converges to object boundaries. This segmentation is treated as a reinforcement learning problem, and scale-space theory is used to achieve robust and efficient multi-scale shape estimation. The learning challenge of end-to-end regression systems can be addressed by learning an iterative strategy to find segmentations.

Description

[0001] related application

[0002] This patent document claims the benefit of the filing date of Provisional US Patent Application Serial No. 62 / 500,604, filed May 3, 2017, under 35 U.S.C. § 119(e), which is hereby incorporated by reference. Background technique

[0003] This embodiment relates to segmentation in medical imaging. Accurate and fast segmentation of anatomical structures is a fundamental task in medical image analysis enabling real-time guidance, quantification, and processing for diagnostic and interventional procedures. Previous solutions for 3D segmentation are based on machine learning-driven active shape models, forward propagation theory, Markov random field methods, or deep image-to-image regression models.

[0004] Active shape models and forward-propagation theoretical solutions propose a parametric surface model that is deformed to fit the boundaries of the target object. Machine learning techniques utilize image databases to learn complex parametric...

Examples

Embodiment Construction

[0014] Multiscale deep reinforcement learning generates multiscale deep reinforcement models for N-dimensional (e.g. 3D) segmentation of objects, where N is an integer greater than 1. In this context, segmentation is formulated as learning an image-driven policy for shape evolution that converges to object boundaries. This segmentation is treated as a reinforcement learning problem, and scale-space theory is used to achieve robust and efficient multi-scale shape estimation. The learning challenge of end-to-end regression systems can be addressed by learning an iterative strategy to find segmentations.

[0015] Although trained as a full segmentation method, the trained policy can be used instead or also for shape refinement as a post-processing step. Any segmentation method provides an initial segmentation. Assuming the original segmentation is used as the initial segmentation in a multi-scale deep reinforcement machine learning model, a machine learning strategy is used to ...

[0001] related application

[0002] This patent document claims the benefit of the filing date of Provisional US Patent Application Serial No. 62 / 500,604, filed May 3, 2017, under 35 U.S.C. § 119(e), which is hereby incorporated by reference. Background technique

[0003] This embodiment relates to segmentation in medical imaging. Accurate and fast segmentation of anatomical structures is a fundamental task in medical image analysis enabling real-time guidance, quantification, and processing for diagnostic and interventional procedures. Previous solutions for 3D segmentation are based on machine learning-driven active shape models, forward propagation theory, Markov random field methods, or deep image-to-image regression models.

[0004] Active shape models and forward-propagation theoretical solutions propose a parametric surface model that is deformed to fit the boundaries of the target object. Machine learning techniques utilize image databases to learn complex parametric...

Examples

Embodiment Construction

[0014] Multiscale deep reinforcement learning generates multiscale deep reinforcement models for N-dimensional (e.g. 3D) segmentation of objects, where N is an integer greater than 1. In this context, segmentation is formulated as learning an image-driven policy for shape evolution that converges to object boundaries. This segmentation is treated as a reinforcement learning problem, and scale-space theory is used to achieve robust and efficient multi-scale shape estimation. The learning challenge of end-to-end regression systems can be addressed by learning an iterative strategy to find segmentations.

[0015] Although trained as a full segmentation method, the trained policy can be used instead or also for shape refinement as a post-processing step. Any segmentation method provides an initial segmentation. Assuming the original segmentation is used as the initial segmentation in a multi-scale deep reinforcement machine learning model, a machine learning strategy is used to ...

Images