NIH Research Festival
–
–
Segmentation of organs and structures in abdominal computed tomography (CT) scans is useful for many clinical applications, such as disease diagnosis and radiotherapy. These CT images are often reconstructed into anisotropic voxelized volumes. This results in coarse and often inconsistent delineations across slices, which are nevertheless used as 3D ground truth for deep segmentation model training and performance reports. Recently, deep self-supervised super-resolution (SSR) has been proposed to improve the quality of CT images without the need for calibrated training data (i.e., low-resolution/high-resolution image pairs). In this work, we investigate whether SSR can be useful to improve the segmentation accuracy of liver in CT scans using deep learning. Results on the public Vin-Dr dataset suggest that SSR can improve segmentation performance not only in terms of volumetric overlap (Dice similarity), but also using more relevant topology preserving evaluation metrics (clDice).
Scientific Focus Area: Computational Biology
This page was last updated on Tuesday, August 6, 2024