Deep LOGISMOS

点击翻译按钮获取中文摘要

Abstract: This is a competitive continuation of a project that already yielded the highly flexible, accurate, and broadly applicable Deep LOGISMOS framework for hybrid deep-learning–graph optimization context- aware n-dimensional medical image segmentation. In this new research phase, we will reformulate the way how Deep LOGISMOS will benefit from topology and shape knowledge learned from training data, how segmentation quality control will increase analysis efficiency, and how an overall performance gain will be reached to always deliver reliable quantitative image analysis that is necessary for adoption in clinical care. To stimulate a new phase of this research project, we hypothesize that: Improvements of relevant information density in training datasets, combined with increased robustness and correctness of object topology and shape estimates, and smart Just-Enough Interaction (JEI) guidance via automated assessment of segmentation quality will universally bring clinically acceptable quantitative analyses in routinely acquired, complex, diagnostic-quality medical images across diverse application areas. The proposed research project brings together multiple major innovations. It is based on a solid scientific premise, builds on already achieved results, extends state-of-the-art of deep learning, and addresses the burning question of efficiently forming compact high-quality training sets of annotated data. Learned object topology and shape patterns will be newly incorporated in Deep LOGISMOS learning models. It improves image analysis correctness and performance in challenging cases, provides new insights in use of automated image segmentation quality assessment to guide the optional JEI processes and increase their efficiency, and further extends the translational and clinical utility and significance of our approach. We will fulfill the following specific aims: 1. Develop novel, robust deep-learning approaches to learn and instill comprehensive topology and shape awareness in Deep LOGISMOS while maintaining its hybrid DL–graph optimization principles. 2. Maximize segmentation success of Deep LOGISMOS + JEI via smart Just-Enough Interaction. 3. Develop efficient approaches for constructing training sets by employing automated quality control. 4. In healthcare-relevant applications, demonstrate that Deep LOGISMOS + JEI leads to high success of fully automated analyses, allows efficient JEI adjudication if needed, and improves segmentation performance in comparison with state-of-the-art segmentation techniques. The next-generation Deep LOGISMOS will bring forth broadly available routine quantification of clinical images, significantly elevating the impact of image-based information in tomorrow’s precision medicine.