Image Descriptors for Weakly Annotated Histopathological Breast Cancer Data.

Area of Science:

• Computer Vision
• Machine Learning
• Digital Pathology

Background:

• Histopathological data analysis relies on machine learning and computer vision for cancerous tissue recognition (CTR).
• Patch-level analysis of high-resolution histopathological images is computationally efficient.
• Current methods often require detailed pixel-level annotations from pathologists.

Purpose of the Study:

• To develop novel feature descriptors for enhanced recognition of malignant regions in histopathological images.
• To augment clinician capabilities within a digital connected health system.
• To reduce the reliance on extensive pixel-level annotations.

Main Methods:

• Introduced the Covariance-Kernel Descriptor (CKD) for compact description of tissue architectures.
• Employed a multiple instance learning framework to derive the Weakly Annotated Image Descriptor (WAID).
• WAID utilizes patch bags with binary labels, eliminating the need for precise tissue delineations.

Main Results:

• CKD achieved 92.83% classification accuracy and 0.98 AUC, outperforming other descriptors on a private breast cancer dataset.
• WAID demonstrated state-of-the-art performance on the BreakHis dataset, with 91.27% and 92.00% correctly classified malignant instances at patient and image levels, respectively.
• Both methods achieved high performance without deep learning schemes.

Conclusions:

• The proposed CKD and WAID offer accurate and efficient tools for histopathological analysis.
• These descriptors can significantly aid medical experts in achieving faster and more precise diagnoses.
• The methods represent an advancement in automated cancerous tissue recognition.