Ensemble based system for whole-slide prostate cancer probability mapping using color texture features.

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Prostate cancer diagnosis relies on histological examination of prostatectomy specimens.
• Accurate segmentation and probability mapping of cancer in histology slides are crucial for clinical decision-making.
• Current methods may face challenges in generalization and data requirements.

Purpose of the Study:

• To develop a robust tile-based computational approach for generating clinically relevant probability maps of prostatic carcinoma.
• To leverage ensemble learning for improved feature selection and classification accuracy.
• To reduce training data requirements while maintaining reliable diagnostic performance.

Main Methods:

• A tile-based methodology was employed for histological sections from radical prostatectomy.
• Ensemble learning, including random forest feature selection and classifier ensembles, was utilized.
• Generalized CIEL*a*b* co-occurrence texture features were selected, and spatial filtering was applied.
• Sample selection strategies with minimal constraints were implemented to reduce training data needs.

Main Results:

• Achieved Area Under the Curve (AUC) values of 95% for cancer probability mapping.
• Demonstrated increased heat-map coherence through spatial filtering of tile-based texture features.
• Validated the effectiveness of ensemble models (random forests, support vector machines) for classification.
• Showcased reduced training data requirements through optimized sample selection.

Conclusions:

• The proposed tile-based approach effectively generates clinically relevant probability maps for prostatic carcinoma.
• Ensemble learning and optimized feature selection enhance accuracy and reduce data dependency.
• The methodology is adaptable to various imaging modalities, features, and histological domains.