CLoDSA: a tool for augmentation in classification, localization, detection, semantic segmentation and instance...

Area of Science:

• Bioimaging
• Computer Vision
• Machine Learning

Background:

• Deep learning in bioimaging requires extensive data, posing challenges due to data scarcity and overfitting.
• Data augmentation, generating synthetic samples, is crucial for improving model performance.
• Existing tools lack support for multi-dimensional image data and advanced tasks like segmentation.

Purpose of the Study:

• To present a versatile strategy for automatic data augmentation of image datasets.
• To address the limitations of current tools for multi-dimensional bioimaging data augmentation.
• To support a wide range of computer vision tasks including classification, localization, detection, and segmentation.

Main Methods:

• Developed a generic augmentation strategy applicable to 2D and multi-dimensional images.
• Implemented the strategy in the open-source package CLoDSA.
• Applied CLoDSA to diverse bioimaging datasets for various analytical tasks.

Main Results:

• CLoDSA effectively augments datasets for classification, localization, detection, semantic segmentation, and instance segmentation.
• Utilized CLoDSA to enhance model accuracy in malaria parasite classification.
• Improved stomata detection and automatic segmentation of neural structures using CLoDSA.

Conclusions:

• CLoDSA is the first library offering image augmentation for multiple computer vision tasks on both 2D and multi-dimensional images.
• The tool addresses a critical gap in bioimaging data augmentation capabilities.
• CLoDSA demonstrates significant improvements in model performance across various bioimaging applications.