Deep Learning for Medical Image Segmentation

TL;DR

This paper compares various deep learning architectures for 3D hippocampal segmentation in MRI, highlighting the trade-offs between performance and computational cost, and demonstrating the effectiveness of a stacked 2D approach.

Contribution

It provides a comparative analysis of 2D, tri-planar, and 3D convolutional architectures for hippocampal segmentation, with insights into their efficiency and accuracy.

Findings

Stacked 2D approach outperforms simple 2D patches.

Tri-planar approach yields better results than 2D with moderate computational increase.

Full 3D architecture offers marginally better accuracy at high computational cost.

Abstract

This report provides an overview of the current state of the art deep learning architectures and optimisation techniques, and uses the ADNI hippocampus MRI dataset as an example to compare the effectiveness and efficiency of different convolutional architectures on the task of patch-based 3-dimensional hippocampal segmentation, which is important in the diagnosis of Alzheimer's Disease. We found that a slightly unconventional "stacked 2D" approach provides much better classification performance than simple 2D patches without requiring significantly more computational power. We also examined the popular "tri-planar" approach used in some recently published studies, and found that it provides much better results than the 2D approaches, but also with a moderate increase in computational power requirement. Finally, we evaluated a full 3D convolutional architecture, and found that it provides marginally better results than the tri-planar approach, but at the cost of a very significant increase in computational power requirement.