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

In biomedical engineering, deep neural networks are commonly used for the diagnosis and assessment of diseases through the interpretation of medical images. The effectiveness of these networks relies heavily on the availability of annotated datasets for training. However, obtaining noise-free and consistent annotations from experts, such as pathologists, radiologists, and biologists, remains a significant challenge. One common task in clinical practice and biological imaging applications is instance segmentation. Though, there is currently a lack of methods and open-source tools for the automated inspection of biomedical instance segmentation datasets concerning noisy annotations. To address this issue, we propose a novel deep learning-based approach for inspecting noisy annotations and provide an accompanying software implementation, AI2Seg, to facilitate its use by domain experts. The performance of the proposed algorithm is demonstrated on the medical MoNuSeg dataset and the biological LIVECell dataset.

Synthesis of large scale 3D microscopic images of 3D cell cultures for training and benchmarking.

The analysis of 3D microscopic cell culture images plays a vital role in the development of new therapeutics. While 3D cell cultures offer a greater similarity to the human organism than adherent cell cultures, they introduce new challenges for automatic evaluation, like increased heterogeneity. Deep learning algorithms are able to outperform conventional analysis methods in such conditions but require a large amount of training data. Due to data size and complexity, the manual annotation of 3D images to generate large datasets is a nearly impossible task. We therefore propose a pipeline that combines conventional simulation methods with deep-learning-based optimization to generate large 3D synthetic images of 3D cell cultures where the labels are known by design. The hybrid procedure helps to keep the generated image structures consistent with the underlying labels. A new approach and an additional measure are introduced to model and evaluate the reduced brightness and quality in deeper image regions. Our analyses show that the deep learning optimization step consistently improves the quality of the generated images. We could also demonstrate that a deep learning segmentation model trained with our synthetic data outperforms a classical segmentation method on real image data. The presented synthesis method allows selecting a segmentation model most suitable for the user's data, providing an ideal basis for further data analysis.

KaIDA: a modular tool for assisting image annotation in deep learning.

Deep learning models achieve high-quality results in image processing. However, to robustly optimize parameters of deep neural networks, large annotated datasets are needed. Image annotation is often performed manually by experts without a comprehensive tool for assistance which is time- consuming, burdensome, and not intuitive. Using the here presented modular Karlsruhe Image Data Annotation (KaIDA) tool, for the first time assisted annotation in various image processing tasks is possible to support users during this process. It aims to simplify annotation, increase user efficiency, enhance annotation quality, and provide additional useful annotation-related functionalities. KaIDA is available open-source at https://git.scc.kit.edu/sc1357/kaida.