Related Experiment Video
Updated: Aug 25, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Data augmentation based on multiple oversampling fusion for medical image segmentation
Liangsheng Wu1,2,3, Jiajun Zhuang2, Weizhao Chen1
1Academy of Interdisciplinary Studies, Guangdong Polytechnic Normal University, Guangzhou, China.
Plos One
|October 18, 2022
Summary
Medical image segmentation models require extensive annotated data, which is challenging to acquire. This study introduces a novel data augmentation technique to improve the segmentation of small lesions, enhancing model performance.
Area of Science:
- Medical Imaging
- Deep Learning
- Computer Vision
Background:
- High-performance medical image segmentation relies on large annotated datasets.
- Acquiring sufficient annotated medical images is difficult.
- Small tissue lesions exacerbate class imbalance in segmentation tasks.
Purpose of the Study:
- To propose a multidimensional data augmentation method for medical image segmentation.
- To address the class imbalance problem caused by small tissue lesions.
- To improve the performance of deep learning models in segmenting small lesions.
Main Methods:
- Implemented a data augmentation strategy combining affine transformation and random oversampling.
- Utilized class weight balancing with a weighted cross-entropy loss function.
- Evaluated four deep neural network models (Mask-RCNN, U-Net, SegNet, DeepLabv3+) on LUNA16 and LiTS17 datasets.
Main Results:
- The proposed data augmentation strategy significantly improved small tissue segmentation performance across all tested architectures.
- Mask-RCNN achieved the best pixelwise segmentation performance for pulmonary nodules (DSC: 0.829) and liver tumors (DSC: 0.879).
- The results are comparable to state-of-the-art methods.
Conclusions:
- The multidimensional data augmentation method effectively enhances medical image segmentation for small lesions.
- Class imbalance in medical image segmentation can be mitigated using weighted loss functions and oversampling techniques.
- The Mask-RCNN model, combined with the proposed augmentation, demonstrates superior performance in segmenting small pulmonary nodules and liver tumors.
