A novel deep learning-based 3D cell segmentation framework for future image-based disease detection.

Area of Science:

• Biomedical imaging
• Computational biology
• Machine learning

Background:

• Accurate cell segmentation is vital for disease diagnosis and treatment.
• Segmenting densely packed cells in 3D images remains a significant challenge.
• Current deep learning methods often require extensive hyperparameter tuning.

Purpose of the Study:

• To develop an automated 3D cell segmentation pipeline (3DCellSeg) that addresses limitations of existing methods.
• To improve the accuracy and efficiency of segmenting densely packed cells in 3D images.
• To provide a robust tool for biomedical and clinical applications like cancer diagnosis.

Main Methods:

• A novel two-stage deep learning pipeline (3DCellSeg) was developed.
• A lightweight convolutional neural network (3DCellSegNet) was designed for efficient voxel-wise mask prediction.
• A custom loss function (3DCellSeg Loss) and clustering algorithm (TASCAN) were introduced to handle clumped and touching cells.

Main Results:

• 3DCellSeg demonstrated superior performance on plant and animal cell datasets, achieving high accuracies (e.g., 95.6% on ATAS).
• The pipeline showed comparable accuracy to baseline methods on the Ovules dataset (82.2%).
• Ablation studies confirmed the contributions of 3DCellSegNet, 3DCellSeg Loss, and TASCAN to overall accuracy and robustness.

Conclusions:

• 3DCellSeg offers a robust and efficient solution for 3D cell segmentation, particularly for challenging dense cell populations.
• The pipeline's performance across diverse datasets highlights its versatility for various cell types and shapes.
• 3DCellSeg has the potential to be a valuable tool in histopathological image analysis for cancer diagnosis and grading.