Extra colloquium Computer Sciences Dr. S. Sengar Singh University of Copenhagen, Denmark

Title: Unet Architecture in Multiplanar Volumetric Segmentation - Validated on 3 Knee MRI Cohorts

Abstract:

The state-of-the-art approaches for knee cartilage segmentation are variants of U-Net/fully convolutional networks and mostly designed for 2D image slices. Even though their success, these models have the following limitations: (1) there is unnecessarily restrictive fusion approach, which force to aggregate only at the same level of encoder and decoder part of network (2) lack of availability of generalized models, those can directly apply over 3D medical images. To overcome, these shortcomings, we propose an MPUNet2+, a multiplanar and UNet2+ based method for knee cartilage segmentation, by (1) reducing the fusion of semantically non-similar features from plain skip connections in UNet (2) generating a multiplanar based generalized model for 3D image with fewer number of parameters. We have implemented MPUNet2+ method on three different real and benchmark Knee MRI image datasets and done extensive quantitative and qualitative evaluation on segmented results. For better generalization purpose, we have tested the performance of multiplanar network with other variants of UNet architectures. It is evident from the results that (1) MPUNet2+ consistently outperforms over other multiplanar based techniques for the task of knee cartilage segmentation across different datasets (2) there are considerably fewer number of parameters and lesser training time as comparison to the tested approaches.