Research on improved DeepLabv3+ image Semantic Segmentation algorithm

Tang, Xiaowei; Xu, Zichen; Xu, Bing

doi:10.1145/3580219.3580244

Cited by 1 publication

(1 citation statement)

References 3 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This study leveraged pre-trained architectures, renowned for their efficacy and efficiency, making it compatible with medical imaging tasks especially when handling limited training datasets. We adopted a dual-stage approach: Stage 1 employed MobileNetV2 [21], capitalizing on its lightweight design for swift feature extraction and LV chamber identification, minimizing computational burden for subsequent stages [22]. Stage 2 utilized ResNet50 [23], to extract finer details necessary for segmenting smaller scars [9], [24]- [26].…”

Section: Network Backbonementioning

confidence: 99%

Automated DeepLabV3+ based model for left ventricle segmentation on short-axis late gadolinium enhancement-magnetic cardiac resonance imaging images

Awang Damit,

Sulaiman,

Osman

et al. 2024

IJECE

View full text Add to dashboard Cite

Accurate segmentation of myocardial scar tissue on late gadolinium enhancement-magnetic cardiac resonance imaging (LGE-CMR) is exceptionally vital for clinical applications, enabling precise diagnosis and effective treatment of various cardiac diseases, such as myocardial infarction and cardiomyopathies. However, the ventricle (LV) variations in the size and shape, artifacts, and image resolution of LGE-CMR has made automatic segmentation of myocardial scar tissue more challenging. While many existing approaches delineate the LV myocardium region using multi-modal segmentation, these models may be computationally complex and suffer from misalignment. Therefore, this study proposed an automatic dual-stage DeepLabV3+ based approach tailored for myocardial scar segmentation on short-axis LGE-MRI exclusively. To segment myocardial scar tissue, the second stage employs the segmented LV chamber from the previous stage. The encoder part of the framework utilizes a MobileNetV2 and ResNet50 backbone for the first and second segmentation, respectively, aiming for optimal resolution of feature maps. Both stages tailor an improved Atrous Spatial Pyramid Pooling module in the DeepLabV3+ model with fine-tuned dilated atrous rates to effectively extract the LV chamber and myocardial scar from the complex LGE-MRI background. Based on the results, the proposed dual-stage network recorded an outstanding segmentation performance, with mean Dice score of 96.02% for LV chamber segmentation and 68.01% for scar tissue extraction.

show abstract