ARCHITECTURAL PATTERNS FOR DIFFERENTIAL DIAGNOSIS OF PROLIFERATIVE BREAST LESIONS FROM HISTOPATHOLOGICAL IMAGES

Area of Science:

• Computational pathology
• Breast cancer diagnostics
• Histopathology image analysis

Background:

• Differentiating benign usual ductal hyperplasia (UDH) from malignant ductal carcinoma in situ (DCIS) poses a diagnostic challenge for pathologists.
• Current computational methods primarily focus on nuclear atypia, overlooking architectural patterns.

Purpose of the Study:

• To develop an improved computational pathology method for classifying proliferative breast lesions.
• To enhance the discrimination between UDH and DCIS by incorporating both cellular morphology and spatial architecture.

Main Methods:

• Utilized a publicly available breast lesion dataset comprising UDH and three grades of DCIS.
• Developed a novel computational approach encoding cellular morphology and spatial architectural patterns.
• Compared performance against state-of-the-art methods focusing on nuclear atypia.

Main Results:

• Achieved a 10% improvement in classification accuracy for discriminating UDH from DCIS compared to existing methods.
• Demonstrated significant improvements in a four-way classification task (UDH and three DCIS grades): 6% accuracy, 8% micro-AUC, and 19% macro-AUC increase.
• The integrated approach of cellular and architectural features outperformed methods relying solely on nuclear atypia.

Conclusions:

• The proposed computational pathology method effectively improves the differential diagnosis of proliferative breast lesions.
• Integrating cellular morphology and spatial architecture offers a more robust approach for classifying UDH and DCIS.
• This advancement holds potential for more accurate and efficient breast cancer diagnostics.