Deep Learning Accurately Quantifies Plasma Cell Percentages on CD138-Stained Bone Marrow Samples.

Area of Science:

• Hematology
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Accurate plasma cell percentage is crucial for diagnosing plasma cell neoplasms.
• Current methods rely on visual estimation from CD138-stained bone marrow biopsies, which lacks precision.
• Deep learning offers a potential solution to enhance diagnostic accuracy.

Purpose of the Study:

• To investigate the efficacy of a deep learning model, specifically a semantic segmentation-based convolutional neural network (CNN), in precisely quantifying plasma cell percentages.
• To compare the CNN's performance against manual pathologist estimations and commercial software.

Main Methods:

• A CNN was trained using pathologist-annotated image patches of CD138+ and CD138- cells from bone marrow biopsies.
• The CNN was validated on an independent dataset using annotations from two pathologists and a non-deep learning commercial software.
• The CNN was scaled to analyze whole slide images (WSIs) and deployed as a web application for practical use.

Main Results:

• The CNN demonstrated high precision, with intraclass correlation coefficients (ICCs) of 0.975 when compared to pathologist #1.
• CNN performance was comparable to pathologist estimations and superior to a non-deep learning commercial software (ICC=0.892).
• Cell-by-cell analysis showed CNN labels were nearly as accurate as pathologist labels.

Conclusions:

• Deep learning, particularly CNNs, can significantly improve the precision of plasma cell percentage quantification in bone marrow diagnostics.
• The developed CNN-based system offers a reliable and accurate tool for assessing plasma cell neoplasms, enhancing diagnostic workflows.
• This technology has the potential to standardize and improve the accuracy of plasma cell neoplasm diagnosis.