Statistical Shape Model for Manifold Regularization: Gleason grading of prostate histology

Area of Science:

• Pathology
• Computer Vision
• Machine Learning

Background:

• Gleason grading of prostate cancer is crucial for determining disease aggressiveness and patient outcomes.
• Manual grading by pathologists can lead to errors and inter-observer variability.
• Automated systems using gland morphology show promise for accurate Gleason pattern classification.

Purpose of the Study:

• To develop a manifold regularization technique using a statistical shape model of manifolds (SSMM) to constrain low-dimensional manifold representations.
• To apply the SSMM for identifying noisy samples and improving out-of-sample extrapolation in automated Gleason grading.
• To enhance the accuracy of distinguishing between Gleason patterns 3 and 4 in prostate histopathology.

Main Methods:

• Implemented a manifold regularization technique constrained by a statistical shape model of manifolds (SSMM).
• Applied the SSMM to identify and remove noisy samples from the manifold.
• Utilized the SSMM for accurate out-of-sample extrapolation of newly acquired gland morphology features.
• Evaluated performance on a prostate histopathology dataset (58 patients) and synthetic datasets.

Main Results:

• Removing noisy samples using SSMM significantly improved the area under the receiver operator characteristic curve (AUC) from 0.779 to 0.832.
• SSMM-based out-of-sample extrapolation also significantly improved AUC to 0.834 compared to 0.779.
• The SSMM demonstrated effectiveness in identifying noisy samples and enhancing classification accuracy.

Conclusions:

• The SSMM is an effective tool for improving the robustness and accuracy of manifold learning in automated Gleason grading.
• This approach can reduce errors associated with manual grading and inter-observer variability.
• The SSMM offers a statistically significant advancement for classifying prostate cancer histopathology.