This is a competitive continuation of a project that already yielded the highly flexible, accurate, and broadly applicable LOGISMOS framework for context-aware n-dimensional image segmentation. To substantially improve and extend its capability, we will develop Deep LOGISMOS that combines and reinforces the complementary advantages of LOGISMOS and deep learning (DL). There is growing need for quantitative failure-free 3D and higher-D image analysis for diagnostic and/or planning purposes. Examples of current use exist in radiation oncology, cardiology, ophthalmology and other areas of routine clinical medicine, many of which however still rely on manual slice-by-slice tracing. This manual nature of such analyses hinders their use in precision medicine. Deep LOGISMOS research proposed here will solve this problem and will offer routine efficient analysis of clinical images of analyzable quality. To stimulate a new phase of this research project, we hypothesize that: Advanced graph-based image segmentation algorithms, when combined with deep-learning-derived application/modality specific parameters and allowing highly efficient expert-analyst guidance working in concert with the segmentation algorithms, will significantly increase quantitative analysis performance in routinely acquired, complex, diagnostic-quality medical images across diverse application areas. The proposed research focuses on establishing an image segmentation and analysis framework combining the strengths of LOGISMOS and DL, developing a new way to efficiently generate training data necessary for learning from examples, forming a failure-free strategy for 3D, 4D, and generally n-D quantitative medical image analysis, and discovering ways for automated segmentation quality control. We will fulfill these specific aims: 1. Develop an efficient approach for building large segmentation training datasets in 3D, 4D, n-D using assisted and suggestive annotations. 2. Develop Deep LOGISMOS, combining two well-established algorithmic strategies ? deep learning and LOGISMOS graph search. 3. Develop and validate methods employing deep learning for quality control of Deep LOGISMOS. 4. In healthcare-relevant applications, demonstrate that Deep LOGISMOS improves segmentation performance in comparison with state-of-the-art segmentation techniques. Deep LOGISMOS will bring broadly available routine quantification of clinical images, positively impacting the role of reliable image-based information in tomorrow?s precision medicine.
Previous phases of this successful research project were devoted to the development of new graph-based methods for multi-surface and/or multi-object multi-dimensional biomedical image segmentation. Deep learning is emerging as an important new way to learn from large sets of examples. This proposal will combine deep learning and graph-based image analysis to maximize their combined strengths and overcome their individual weaknesses with the overarching goal to facilitate routine use of quantitative medical image analysis techniques in personalized medical care.
Showing the most recent 10 out of 99 publications
