Computer Vision News 42 Congrats, Doctor Anne-Marie! Medical Image Segmentation is an important task in image analysis pipelines, serving as a prerequisite for many downstream applications. While some may consider segmentation a solved task, I believe there are still challenges in segmentation that are interesting to work on. Here I highlight some methods we worked on in the last years. Organ hallucinations: We introduced HALOS at IPMI 2023, addressing a unique challenge – organ hallucinations. When investigating the performance of usually wellperforming models like nnU-Net on large-scale and diverse datasets like the UK-Biobank, we found that these models struggle when they are confronted with anatomical changes post-organ resection surgery, such as cholecystectomy or nephrectomy. The models, interestingly, tend to segment organs that no longer exist. We termed this phenomenon organ hallucination. HALOS tackles this by integrating an organ existence classifier and using feature fusion modules to enhance the segmentation process with information about organ existence. A standout feature of HALOS is its flexibility: it can utilize ground truth labels for organ existence when available, or alternatively, rely on the classifier's output. This approach shows promise for extension to other organ resection cases and might be useful in other scenarios where anatomy significantly differs from the majority of training data scans. Cortical Surface Reconstruction: A large part of my research focused on cortical surface reconstruction. Anne-Marie Rickmann recently completed her doctoral degree at the LudwigMaximilians University Munich (LMU) under the supervision of Prof. Christian Wachinger. Her research focused on developing deep learning techniques for medical image segmentation of 3D MRI and CT scans. Anne will continue her research as a postdoctoral researcher at Yale School of Medicine in the Radiology and Biomedical Imaging department, where she will work with James Duncan and Albert Sinusas on developing deep learning methods for multimodal cardiac image analysis. Congrats Doctor Anne-Marie!
RkJQdWJsaXNoZXIy NTc3NzU=