BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Jinming Duan (University of Birmingham) DTSTART:20200714T130000Z DTEND:20200714T140000Z DTSTAMP:20260423T034449Z UID:DSCSS/2 DESCRIPTION:Title: Ca rdiac Magnetic Resonance Image Segmentation with Anatomical Knowledge\ nby Jinming Duan (University of Birmingham) as part of Data Science and Co mputational Statistics Seminar\n\n\nAbstract\nThis talk focuses on segment ation of cardiac magnetic resonance (CMR) images from both healthy and pat hological subjects. Specifically\, we will propose three different approac hes that explicitly consider geometry (anatomy) information of the heart.\ n\nFirst\, we introduce a novel deep level set method\, which explicitly c onsiders the image features learned from a deep neural network. To this en d\, we estimate joint probability maps over both region and edge locations in CMR images using a fully convolutional network. Due to the distinct mo rphology of pulmonary hypertension (PH) hearts\, these probability maps ca n then be incorporated in a single nested level set optimisation framework to achieve multi-region segmentation with high efficiency. We show result s on CMR cine images and demonstrate that the proposed method leads to sub stantial improvements for CMR image segmentation in PH patients.\n\nSecond \, we propose a multi-task deep learning approach with atlas propagation t o develop a shape-refined bi-ventricular segmentation pipeline for short-a xis CMR volumetric images. The pipeline combines the computational advanta ge of 2.5D FCNs networks and the capability of addressing 3D spatial consi stency without compromising segmentation accuracy. A refinement step is in troduced for overcoming image artefacts (e.g.\, due to different breath-ho ld positions and large slice thickness)\, which preclude the creation of a natomically meaningful 3D cardiac shapes. Extensive numerical experiments on the two large datasets show that our method is robust and capable of pr oducing accurate\, high-resolution\, and anatomically smooth bi-ventricula r 3D models\, despite the presence of artefacts in input CMR volumes.\n\nL astly\, accelerating the CMR acquisition is essential. However\, reconstru cting high-quality images from accelerated CMR acquisition is a nontrivial problem. As such\, I will show how deep neural networks can be developed to bypass the usual image reconstruction stage. The method applies shape p rior knowledge through an auto-encoder. Due to the prior knowledge\, we im proved both the CMR acquisition time and segmentation accuracy.\n LOCATION:https://researchseminars.org/talk/DSCSS/2/ END:VEVENT END:VCALENDAR