[论文解读] DeepGI: An Automated Approach for Gastrointestinal Tract Segmentation in MRI Scans

Ye Zhang, Yulu Gong+3 more|arXiv (Cornell University)|Jan 27, 2024

Radiomics and Machine Learning in Medical Imaging被引用 13

一句话总结

本论文提出一个 tri-path 深度学习模型（Inception-V4 classification、2.5D U-Net++ with VGG19、以及 Grayscale Edge U-Net）用于 MRI 中 GI 道段的放疗计划自动分割，采用集成预测和专门的预处理。

ABSTRACT

Gastrointestinal (GI) tract cancers pose a global health challenge, demanding precise radiotherapy planning for optimal treatment outcomes. This paper introduces a cutting-edge approach to automate the segmentation of GI tract regions in magnetic resonance imaging (MRI) scans. Leveraging advanced deep learning architectures, the proposed model integrates Inception-V4 for initial classification, UNet++ with a VGG19 encoder for 2.5D data, and Edge UNet for grayscale data segmentation. Meticulous data preprocessing, including innovative 2.5D processing, is employed to enhance adaptability, robustness, and accuracy. This work addresses the manual and time-consuming segmentation process in current radiotherapy planning, presenting a unified model that captures intricate anatomical details. The integration of diverse architectures, each specializing in unique aspects of the segmentation task, signifies a novel and comprehensive solution. This model emerges as an efficient and accurate tool for clinicians, marking a significant advancement in the field of GI tract image segmentation for radiotherapy planning.

研究动机与目标

• Reduce manual delineation in radiotherapy planning by automating GI tract segmentation in MRI scans.
• Develop a unified model combining multiple specialized architectures to capture diverse anatomical details.
• Improve robustness and accuracy through innovative data preprocessing (2.5D and grayscale processing).
• Provide an efficient tool to standardize GI tract delineation and reduce inter-observer variability.

提出的方法

• A tri-path architecture combines Inception-V4 for initial healthy-organ classification, 2.5D U-Net++ with a VGG19 encoder for detailed segmentation, and Grayscale Edge U-Net with HED-based edge features.
• Data preprocessing includes spatial augmentation (resize to 320x384, horizontal flip, rotation, elastic distortion, coarse dropout) and grayscale intensity augmentation; 2.5D processing stacks three consecutive MRI slices to create 2.5D inputs.
• Edges are incorporated via Edge U-Net with MBConv blocks and HED edge maps; inferences from all three pathways are averaged to form final segmentation.
• Evaluation uses Dice Coefficient and 3D Hausdorff Distance, combined as a weighted composite score (Score = 0.4*Dice + 0.6*3D Hausdorff).
• The model targets segmentation of colon, small intestine, and stomach regions in GI tract MRI scans.

实验结果

研究问题

• RQ1 How well do grayscale, 2.5D, and edge-aware pathways perform individually for GI tract segmentation in MRI?
• RQ2 Does integrating Inception-V4, 2.5D U-Net++ with VGG19, and Edge U-Net via averaging improve segmentation accuracy over individual paths?
• RQ3 Which encoder architectures best suit grayscale vs. 2.5D MRI data for GI segmentation?
• RQ4 What preprocessing strategies most effectively enhance model robustness across imaging scenarios?

主要发现

模型	编码器	验证分数
UNet	ResNet50	0.71599
UNet	Inception-V4	0.71002
UNet	Xception	0.73761
UNet	EfficientNet-B0	0.68033
UNet	VGG19	0.78925
UNet++	ResNet50	0.7899
UNet++	Inception-V4	0.80095
UNet++	Xception	0.79711
UNet++	EfficientNet-B0	0.71372
UNet++	VGG19	0.80717
Edge UNet	-	0.84046
UNet++	ResNet50	0.80138
UNet++	Xception	0.7961
UNet++	VGG19	0.84984

• Edge U-Net achieves the strongest performance for grayscale image segmentation with a validation score of 0.84046.
• For 2.5D data, UNet++ with a VGG19 encoder yields the best validation score of 0.84984.
• Across grayscale experiments, UNet++ with VGG19 also performs strongly with 0.80717, while UNet with VGG19 reaches 0.78925.
• Overall, grayscale Edge U-Net outperforms others in grayscale, while 2.5D UNet++ with VGG19 leads in 2.5D segmentation.
• The proposed ensemble approach combines outputs from Inception-V4, 2.5D U-Net++, and Grayscale Edge U-Net to enhance segmentation robustness.