Multi-Organ Gland Segmentation Using Deep Learning.

Area of Science:

• Computational pathology
• Digital pathology
• Machine learning in oncology

Background:

• Clinical morphological analysis of histopathology is crucial for cancer diagnosis.
• Computational pathology offers automated, objective, and scalable analysis of histopathology samples.
• Gland segmentation is vital for cancer diagnosis but challenging due to morphological variability and the need for extensive organ-specific annotations.

Purpose of the Study:

• To investigate cross-domain approximation for gland segmentation, reducing reliance on organ-specific annotations.
• To develop and evaluate a deep learning model for stroma segmentation as a pathway to generalize gland segmentation across different organs.
• To compare the performance of gland segmentation versus stroma segmentation in a cross-organ context.

Main Methods:

• Two Dense-U-Net models were developed and trained on H&E stained colon adenocarcinoma samples for gland and stroma segmentation.
• The models were evaluated on independent colon adenocarcinoma and breast invasive cancer datasets.
• Performance was assessed using Dice coefficient and Hausdorff distance metrics.

Main Results:

• The stroma segmentation network performed comparably to the gland segmentation network on the colon dataset.
• On the breast cancer dataset, the stroma segmentation approach significantly outperformed gland segmentation, demonstrating superior generalization.
• The findings highlight the potential of stroma segmentation for multi-organ gland segmentation.

Conclusions:

• Segmenting stromal components offers better generalization across different organ types compared to direct gland segmentation.
• This stroma-centric approach can reduce the need for extensive, organ-specific annotations in computational pathology.
• The proposed method shows promise for developing more extensible and scalable deep learning tools for automated cancer diagnosis.