Automated mitosis detection in histopathology using morphological and multi-channel statistics features

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Mitosis counting is vital for cancer diagnosis and grading per the Nottingham system.
• Manual mitosis counting suffers from significant reader variability, impacting diagnostic accuracy.

Purpose of the Study:

• To enhance mitosis detection accuracy by optimizing color channel selection for statistical and morphological feature extraction.
• To develop a computational framework assisting pathologists in precise mitosis identification.

Main Methods:

• A framework combining statistical and morphological feature analysis across selected color channels was developed.
• Candidate mitosis regions were detected using Laplacian of Gaussian, thresholding, morphology, and active contour models on blue-ratio images.
• A decision tree classifier was employed, utilizing 143 features (morphological, statistical, texture) extracted from candidate regions.

Main Results:

• The method achieved a 74% detection rate and 70% precision on Aperio images from the MITOS dataset.
• On Hamamatsu images, the system reached a 71% detection rate and 56% precision.
• F-measure scores of 72% (Aperio) and 63% (Hamamatsu) demonstrate robust performance.

Conclusions:

• The proposed multi-channel feature computation scheme effectively captures statistical features for mitosis detection.
• The robust model proved highly efficient in the MITOS international benchmark for cancer diagnosis.
• Future work includes exploring color deconvolution and advanced candidate detection techniques.