Deep Models for Stroke Segmentation: Do Complex Architectures Always Perform Better?

Ahmed M.S. Soliman

doi:10.1109/ACCESS.2024.3522214

Deep Models for Stroke Segmentation: Do Complex Architectures Always Perform Better?

2024-01-01

0

OA

AI

DOI：10.1109/ACCESS.2024.3522214

原文链接

原文求助

摘要

En 中文

The accurate segmentation of stroke lesions is crucial for the diagnosis and treatment of stroke patients, as it provides spatial information about affected brain regions and the extent of damage. While conventional manual techniques are time-consuming and prone to errors, advanced deep learning models have shown promising results in medical image segmentation. Recently, several complex architectures, such as vision Transformers and attention-based convolutional neural networks (CNNs), have been introduced for this task. However, the question remains whether such high-level designs are necessary to achieve the best results for all segmentation cases. In this paper, we evaluated the performance of four types of deep models for stroke segmentation: 1) a pure Transformer-based architecture (DAE-Former), 2) two advanced CNN-based models (LKA and DLKA) with attention mechanisms, 3) a hybrid model that incorporates CNNs with Transformers (FCT), and 4) the well-known self-adaptive nnU-Net framework. We examined their performance on two publicly available datasets, ISLES 2022 and ATLAS v2.0, and found that the nnU-Net, with its relatively simple design, achieved the best results among all the models tested. Furthermore, we investigated the impact of an imbalanced distribution of the number of unconnected components in each slice, as a representation of common variability in stroke segmentation. Our findings reveal a potential robustness issue of Transformers to such variability, which may explain their unexpected weak performance. Additionally, the success of nnU-Net underscores the significant impact of pre- and post-processing techniques in enhancing segmentation results, rather than solely focusing on architectural designs. These findings suggest that proposed complex architectures may be task-specific and simpler models with appropriate pre-/post-processing pipeline can be equally or more effective in generalization across different tasks in medical image segmentation.

Keyword:

Image segmentation

Transformers

Computer architecture

Stroke (medical condition)

Feature extraction

Data models

Computer vision

Decoding

Convolutional neural networks

Accuracy

deep learning

nnU-Net

stroke segmentation

vision Transformer

Deep Models for Stroke Segmentation: Do Complex Architectures Always Perform Better?

摘要

期刊

机构

引用论文