Direct Cellularity Estimation on Breast Cancer Histopathology Images Using Transfer Learning

Area of Science:

• Computational pathology
• Digital pathology
• Breast cancer research

Background:

• Residual cancer burden (RCB) assessment is crucial for evaluating post-neoadjuvant breast cancer response.
• Manual estimation of cancer cellularity from H&E-stained slides is a time-consuming and subjective bottleneck in RCB workflow.
• Accurate cellularity estimation is vital for reliable RCB scoring and treatment response prediction.

Purpose of the Study:

• To develop an automated, direct method for estimating cancer cellularity from histopathological image patches.
• To improve the accuracy and efficiency of cellularity assessment in breast cancer analysis.
• To provide a tool that reduces reliance on manual nucleus segmentation and classification.

Main Methods:

• Utilized deep feature representation, tree boosting, and support vector machine (SVM) algorithms.
• Developed a method to estimate cellularity directly from image patches, bypassing nucleus segmentation.
• Trained and validated the model on a dataset of 2394 training and 185 testing image patches.

Main Results:

• The automated method demonstrated strong correlations with human pathologist estimations.
• Achieved high intraclass correlation (ICC) of 0.94, Kendall's tau of 0.83, and prediction probability of 0.93.
• Outperformed two other comparative methods in accuracy and reliability.

Conclusions:

• The developed automated method accurately estimates cancer cellularity in histopathological images.
• This approach offers a more efficient and objective alternative to manual cellularity assessment for RCB.
• The method enhances the accuracy of breast cancer response evaluation without requiring individual nucleus annotations.