MICCAI 2022 Daily – Monday

13 DAILY MICCAI Monday Workshop: STACOM The second challenge ( CMRxMotion ) was organized by yet another participant from Fudan University, Shuo Wang and hosted by Chen Chen and Ouyang Cheng from Imperial College. It addressed respiratory motion artefacts with two main tasks: 1) CMR image quality assessment and 2) robust CMR segmentation. This session included talks on a wide range of techniques- from recurrent neural networks and “insane” data augmentation to deformable convolutions, multi-task learning and ensemble classification frameworks. Yasmina Al Khalil started introducing the method OPENGTN, which won 3 rd Place on both tasks, and included two sections, an auto- encoder trained to reconstruct images with noise for prediction of quality control, and then an ensemble of models to improve robustness through data augmentation helped by region-based training which segmented apical, middle, and basal slices separately. The 2 nd place on Task 1 was won by the Philips CTS method from Xiuzheng Yue , combining deep learning for global view and machine learning for LV radiomics feature extraction through voting, while the 1 st place was achieved by UON_IMA, from Ruizhe LI , where the author used a biased voting strategy to aggregate the decisions from different patch-based models. The last session covered some echo-based projects such as Unsupervised Echocardiography Registration through Patch-based MLPs and Transformers, a patch-based MLP/Transformer method to extract features from echo data, and another one by Matthias Ivantsits (image next page) who built an end-to-end oriented mitral valve DL surface reconstruction, applied to 3D TEE data and employing a Voxel- Encoder, Voxel-Decoder and a Mesh-Decoder.

RkJQdWJsaXNoZXIy NTc3NzU=