ncku-talk2022-05-19

理學論壇暨數學系專題演講

DATE 2022-05-19　15:10-17:00

PLACE 理學院教學大樓36102教室

SPEAKER 林文偉教授 (國家講座)（國立陽明交通大學應用數學系）

TITLE Computational Conformal Geometry with Optimal Mass Transportations and its Application on 3D Brain Tumor Segmentations
線上專題演講

ABSTRACT 線上專題演講
In this talk, we would like to introduce the computational conformal geometry with optimal mass transportation techniques and its applications on 3D medical image detections and segmentations. The well-known uniformization theorem shows that a closed surface of genus-zero is equivalently conformal to a unit sphere. However, the numerical method and its convergence should be addressed. We will propose efficient algorithms on conformal energy minimization (CEM), stretch energy minimization (SEM) and volume stretch energy minimization (VSEM) for finding the conformal (angle-preserving) and equiareal (area-preserving) parametrizations, respectively, between a simply connected closed surface and a sphere, as well as, the volume-preserving parametrization between a 3-manifold with a genus-zero boundary and a unit ball. Based on the SEM and VSEM algorithms we further develop the reliable and robust algorithms for solving the optimal mass transportation (OMT) between an irregular 3D domain and a unit ball, while minimizing the deformation cost, and keeping the minimal distortion and the local mass ratios unchanged. Combining the proposed OMT with the U-net machine learning algorithm, we develop a novel two-phase OMT algorithm successfully applying for the detection and segmentation of 3D brain tumors with high training and validation Dice scores. For training, good Dice scores: 0.9538 for the WT (whole tumor), 0.9546 for the TC (tumor core) and 0.9093 for the ET (enhanced tumor) can be obtained. For validation, the Dice scores of WT, TC and ET with mesh refinement and ensemble voting postprocessing can reach 0.9371, 0.9062 and 0.8747. A significant accuracy improvement in brain tumor detection and segmentation is achieved. Furthermore, It takes within 200 seconds to complete the whole brain tumor segmentation process for each new brain sample.
會議鏈結: https://nckucc.webex.com/nckucc/j.php?MTID=md7b8928b1dca3f8922c5c64c288418fc
會議號: n8UGpnGC4J3

理學論壇暨數學系專題演講

DATE	2022-05-19　15:10-17:00

PLACE	理學院教學大樓36102教室

SPEAKER	林文偉教授 (國家講座)（國立陽明交通大學應用數學系）

TITLE	Computational Conformal Geometry with Optimal Mass Transportations and its Application on 3D Brain Tumor Segmentations 線上專題演講

ABSTRACT	線上專題演講 In this talk, we would like to introduce the computational conformal geometry with optimal mass transportation techniques and its applications on 3D medical image detections and segmentations. The well-known uniformization theorem shows that a closed surface of genus-zero is equivalently conformal to a unit sphere. However, the numerical method and its convergence should be addressed. We will propose efficient algorithms on conformal energy minimization (CEM), stretch energy minimization (SEM) and volume stretch energy minimization (VSEM) for finding the conformal (angle-preserving) and equiareal (area-preserving) parametrizations, respectively, between a simply connected closed surface and a sphere, as well as, the volume-preserving parametrization between a 3-manifold with a genus-zero boundary and a unit ball. Based on the SEM and VSEM algorithms we further develop the reliable and robust algorithms for solving the optimal mass transportation (OMT) between an irregular 3D domain and a unit ball, while minimizing the deformation cost, and keeping the minimal distortion and the local mass ratios unchanged. Combining the proposed OMT with the U-net machine learning algorithm, we develop a novel two-phase OMT algorithm successfully applying for the detection and segmentation of 3D brain tumors with high training and validation Dice scores. For training, good Dice scores: 0.9538 for the WT (whole tumor), 0.9546 for the TC (tumor core) and 0.9093 for the ET (enhanced tumor) can be obtained. For validation, the Dice scores of WT, TC and ET with mesh refinement and ensemble voting postprocessing can reach 0.9371, 0.9062 and 0.8747. A significant accuracy improvement in brain tumor detection and segmentation is achieved. Furthermore, It takes within 200 seconds to complete the whole brain tumor segmentation process for each new brain sample. 會議鏈結: https://nckucc.webex.com/nckucc/j.php?MTID=md7b8928b1dca3f8922c5c64c288418fc 會議號: n8UGpnGC4J3