Deep information-guided feature refinement network for colorectal gland segmentation.

Area of Science:

• Computational pathology
• Digital image analysis
• Histopathology

Background:

• Accurate quantification of colorectal histopathological images relies on precise gland segmentation.
• Challenges include varying glandular morphology, similar appearances of malignant glands and non-gland tissues, and difficulties in segmenting densely packed glands.

Purpose of the Study:

• To develop an improved deep learning network for precise gland segmentation in colorectal histopathology.
• To address the challenges of varying morphology and dense packing of glands for enhanced diagnostic accuracy.

Main Methods:

• A deep information-guided feature refinement network was proposed.
• Key components include a deepened backbone for effective feature extraction, a Multi-Scale Fusion module to expand the receptive field, and a Multi-Scale Edge-Refined module to strengthen gland boundaries.

Main Results:

• The proposed network demonstrated superior performance compared to eight other deep learning methods on Test B.
• Achieved an F1 score of 0.917 (Test A) and 0.876 (Test B), object-level Dice of 0.921 (Test A) and 0.884 (Test B), and object-level Hausdorff of 43.428 (Test A) and 87.132 (Test B).

Conclusions:

• The network effectively extracts features with high representational ability and enhances edge details for improved segmentation.
• Significantly enhances segmentation performance on malignant glands and yields better results for multi-scale and tightly clustered glands.