AI²Seg: A Method and Tool for AI-based Annotation Inspection of Biomedical Instance Segmentation Datasets

Area of Science:

• Biomedical engineering
• Medical image analysis
• Artificial intelligence in healthcare

Background:

• Deep neural networks (DNNs) are crucial for medical image analysis and disease diagnosis.
• High-quality annotated datasets are essential for training DNNs, but expert annotations can be noisy and inconsistent.
• Instance segmentation is a key task in biomedical imaging, yet automated tools for inspecting annotation quality are lacking.

Purpose of the Study:

• To develop a novel deep learning-based approach for automated inspection of noisy annotations in biomedical instance segmentation datasets.
• To provide an open-source software implementation (AI²Seg) for domain experts to easily use the proposed method.

Main Methods:

• A deep learning-based algorithm was developed to identify and inspect noisy annotations within biomedical instance segmentation datasets.
• The method was implemented as an open-source software tool named AI²Seg.
• The performance was evaluated using two distinct datasets: the medical MoNuSeg dataset and the biological LIVECell dataset.

Main Results:

• The proposed deep learning approach effectively identifies noisy annotations in biomedical instance segmentation datasets.
• The AI²Seg software facilitates the practical application of this inspection method by domain experts.
• Demonstrated performance on both medical (MoNuSeg) and biological (LIVECell) imaging datasets.

Conclusions:

• The developed deep learning method and AI²Seg software address the critical need for automated inspection of noisy annotations in biomedical imaging.
• This work contributes to improving the reliability and accuracy of AI models trained on biomedical image data.
• Facilitating high-quality data annotation is essential for advancing AI applications in healthcare and biological research.