Test-time augmentation for deep learning-based cell segmentation on microscopy images.

Area of Science:

• Computational Biology
• Machine Learning
• Image Analysis

Background:

• Deep learning models require extensive annotated data for high accuracy in microscopy image processing.
• Test-time augmentation (TTA) offers an alternative to training set expansion for improving model performance.

Purpose of the Study:

• To integrate TTA into U-Net (semantic segmentation) and Mask R-CNN (instance segmentation) for single-cell microscopy image analysis.
• To evaluate the impact of simple TTA techniques on prediction accuracy.

Main Methods:

• Implemented TTA by generating transformed image versions at prediction time and merging results.
• Applied TTA to semantic segmentation (U-Net) and instance segmentation (Mask R-CNN) models.
• Utilized microscopy images from the Data Science Bowl 2018 and other cell culture datasets.

Main Results:

• TTA significantly improved prediction accuracy for both U-Net and Mask R-CNN models, even with basic augmentations like rotation and flipping.
• The highest-scoring method from the Data Science Bowl 2018 was further enhanced using TTA.
• The TTA-enhanced method achieved an all-time best score in the Data Science Bowl nuclei segmentation competition.

Conclusions:

• TTA is an effective strategy for boosting the accuracy of deep learning models in single-cell microscopy image segmentation.
• Simple TTA methods, when properly merged, can lead to substantial improvements in prediction performance.
• The proposed TTA approach demonstrates state-of-the-art performance in nuclei segmentation tasks.