CNN cascades for segmenting sparse objects in gigapixel whole slide images.

Area of Science:

• Digital pathology
• Computational image analysis
• Histopathology

Background:

• Digitization of histopathological tissue generates large datasets.
• Need for computer-based image analysis systems for large gigapixel images.
• Challenges in analyzing sparse, small objects-of-interest in histopathology.

Purpose of the Study:

• To develop and evaluate computer-based image analysis approaches for segmenting small objects in gigapixel histopathology images.
• To propose and compare two Convolutional Neural Network (CNN) cascade approaches for glomeruli segmentation.
• To assess the performance against conventional fully-convolutional networks.

Main Methods:

• Development of two CNN cascade approaches for image segmentation.
• Application of CNN cascades to segment glomeruli in renal whole slide images.
• Comparison with fully-convolutional networks using eight-fold cross-validation.
• Reporting of pixel-level Dice similarity coefficient, precision, and recall.

Main Results:

• The best performing CNN cascade approach outperformed single CNNs.
• Achieved a pixel-level Dice similarity coefficient of 0.90 (precision: 0.89, recall: 0.92).
• Demonstrated excellent segmentation accuracy and low computing time compared to previous methods.

Conclusions:

• The proposed CNN cascade network is a powerful tool for accurate automated segmentation of renal whole slide images.
• Facilitates fully-automated big data analyses for medical treatment assessment.
• The approach is adaptable to other biomedical image analysis scenarios.