Source-linked AI summary
X-Net: Brain Stroke Lesion Segmentation Based on Depthwise Separable Convolution and Long-range Dependencies
Kehan Qi, Hao Yang, Cheng Li, Zaiyi Liu, Meiyun Wang, Qiegen Liu, Shanshan Wang
TL;DR
Brain stroke lesion segmentation is needed for clinical measurement and treatment planning, but existing approaches can be parameter-heavy and limited in capturing long-range dependencies. The paper proposes X-Net, combining depthwise separable convolution with a Feature Similarity Module for compact, dense contextual modeling. X-Net achieves better performance than six state-of-the-art methods on ATLAS while using 15.1M trainable parameters.
Problem
Automatic brain stroke lesion segmentation is clinically needed, while existing approaches can use many parameters and inadequately capture long-range dependencies.
Method
X-Net combines depthwise separable convolution for parameter reduction with a non-local Feature Similarity Module for extracting long-range dependencies and dense contextual information.
Results
X-Net performs better than six state-of-the-art methods on the ATLAS dataset and has 15.1M trainable parameters.
Takeaways & Limitations
The proposed model provides an automated brain stroke lesion segmentation approach with fewer trainable parameters and effective long-range context extraction.
Abstract
from arXiv · showhide
The morbidity of brain stroke increased rapidly in the past few years. To help specialists in lesion measurements and treatment planning, automatic segmentation methods are critically required for clinical practices. Recently, approaches based on deep learning and methods for contextual information extraction have served in many image segmentation tasks. However, their performances are limited due to the insufficient training of a large number of parameters, which sometimes fail in capturing long-range dependencies. To address these issues, we propose a depthwise separable convolution based X-Net that designs a nonlocal operation namely Feature Similarity Module (FSM) to capture long-range dependencies. The adopted depthwise convolution allows to reduce the network size, while the developed FSM provides a more effective, dense contextual information extraction and thus facilitates better segmentation. The effectiveness of X-Net was evaluated on an open dataset Anatomical Tracings of Lesions After Stroke (ATLAS) with superior performance achieved compared to other six state-of-the-art approaches. We make our code and models available at https://github.com/Andrewsher/X-Net.
1 Introduction
Brain stroke lesion segmentation is clinically important but difficult because manual annotation is time-consuming, lesions vary substantially, and boundaries can be fuzzy. X-Net addresses these challenges with fewer parameters and long-range context modeling, achieving better results than six state-of-the-art methods on ATLAS.
- Manual slice-by-slice lesion segmentation is time-consuming and relies heavily on radiologists’ subjective perceptions, motivating automatic clinical methods.
- Lesion shape, scale, size, location, and boundary clarity vary, making accurate automatic segmentation challenging.
- X-Net replaces classical convolution with depthwise separable convolution to reduce trainable parameters.
- The Feature Similarity Module is a non-local operation that extracts long-range dependencies and dense contextual information for lesion segmentation.
- X-Net achieves better brain stroke lesion segmentation results than six state-of-the-art methods on the open ATLAS dataset.
2 Method
X-Net combines depthwise separable convolution in X-blocks with a Feature Similarity Module to reduce parameters and capture long-range contextual dependencies for stroke-lesion segmentation.
- Feature Similarity Module: The FSM first projects the input feature map, computes pairwise position relations, and combines the resulting contextual feature with the original feature.The input is filtered through a convolution, while the value representation is generated by a 1 × 1 convolution before relation-weighted aggregation.
- Feature Similarity Module: FSM computes relationships between all feature-map positions, aggregates transformed features with the original representation, and applies a residual connection.The module uses embedded features and dot-product softmax relations to construct a context-enriched feature map.
- X-block: X-blocks use cascaded depthwise separable convolutions and residual connections to reduce trainable parameters while preserving feature extraction.Each X-block contains three cascaded 3 × 3 separable-convolution layers, with a 1 × 1 convolution in the residual path.
- X-Net architecture: X-Net integrates X-blocks, FSM, an encoder-decoder architecture, and skip connections for brain stroke lesion segmentation.X-blocks and pooling form the encoder, while FSM extracts long-range context and the decoder restores spatial resolution.
3 Experimental Results
X-Net was evaluated on the ATLAS dataset using cross-validation, ablation studies, and comparisons with six state-of-the-art methods. It achieved stronger segmentation results while using 15.1M trainable parameters.
- Experimental Setup: The experiments were conducted on ATLAS, an open-source dataset containing specialist-manually segmented stroke lesions.
- Experimental Setup: The evaluation used 5-fold cross-validation with Dice, IoU, precision, and recall averaged across validation images.ATLAS contains 229 normalized 3D T1-weighted MR images from 11 cohorts, totaling 43,281 2D slices.
- Ablation Analysis of Feature Similarity Module: FSM improved Dice, IoU, and recall over base models, although precision decreased slightly.The authors consider the recall gain worthwhile because detecting all stroke lesions is important.
- Ablation Analysis of Feature Similarity Module: FSM was more effective in U-Net and ResUNet than in X-Net, suggesting X-Net’s X-block already captured some interdependencies.
- Comparison with State-of-the-art Methods: 0.0126, 0.0164, and 0.0006 improvements on Dice, IoU, and precision respectively were achieved over six state-of-the-art methods on ATLAS.Table 2 reports the comparison, while the study also notes 15.1M trainable parameters.
4 Conclusion
The paper concludes that X-Net combines parameter-efficient convolution with dense contextual modeling for brain stroke lesion segmentation. Experiments on ATLAS showed better performance than existing models.
- 4 Conclusion: X-Net achieved better performance than existing models in experiments on the ATLAS dataset.
- 4 Conclusion: Depthwise separable convolution reduces trainable parameters, while FSM probes dense contextual information through long-range dependencies.
- 4 Conclusion: The proposed model addresses large parameter counts and inefficient long-range context capture in existing approaches.