Label-Preserving Data Augmentation for Robust Segmentation of Thin Structure in MRI
- Author:
Wooseung KIM
1
;
Yeonah KANG
;
Seokhwan LEE
;
Ho-Joon LEE
;
Yoonho NAM
Author Information
- Publication Type:Original Article
- From:Investigative Magnetic Resonance Imaging 2024;28(3):107-113
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:This study aims to enhance the performance of deep learning models for segmenting thin anatomical structures in medical images by introducing a label-preserving data-augmentation strategy.
Materials and Methods:We developed a data-augmentation technique that applies geometric transformations and their inverses sequentially to input images while preserving the corresponding labels. This method was evaluated on inner ear magnetic resonance images for the automatic segmentation of semicircular canals characterized by thin and circular structures. The dataset included both internal and external samples. For the internal dataset, 70 subjects were used for model training and eight subjects for internal validation. Images were acquired using a 3 tesla magnetic resonance imaging scanner with a three-dimensional high-resolution T2 sequence, and ground-truth segmentations were manually annotated by an experienced radiologist. For external validation, four subjects from a public dataset (Vestibular-Schwannoma-SEG dataset, part of The Cancer Imaging Archive) with high-resolution T2 images for inner ear analysis were used. We performed quantitative evaluations using metrics such as Dice, intersection over union (IoU), 95% Hausdorff distance (HD), and average surface distance (ASD). A qualitative visual assessment was also performed.
Results:The proposed model exhibited improved performance in semicircular canal segmentation in both quantitative and qualitative evaluations. Metrics such as Dice, IoU, 95% HD, and ASD indicated better performance than conventional methods.
Conclusion:The proposed label-preserving data augmentation method improves the segmentation of thin anatomical structures in medical images and offers a robust and efficient solution for enhancing deep learning models in medical imaging.
