VFM-SSL-BMADCC-Framework: vision foundation model and self-supervised learning based automated framework for differential cell counts on whole-slide bone marrow aspirate smears.

Background: Differential cell counts (DCCs) on bone marrow aspirate (BMA) smear is a critical step in the diagnosis and treatment of blood and bone marrow diseases. However, manual counts relies on the experience of pathologists and is very time-consuming. In recent years, deep learning-based intelligent cell detection models have achieved high detection accuracy on datasets of specific diseases and medical centers, but these models depend on a large amount of annotated data and have poor generalization. When the detection task changes or model is applied in different medical centers, we need to re-annotate a large amount of data and retrain the model to ensure detection accuracy.

Methods: To address the above issues, we designed an automated framework for whole-slide bone marrow aspirate smear differential cell counts (BMADCC), called VFM-SSL-BMADCC-Framework. This framework only requires whole-slide images (WSIs) as input to generate DCCs. The vision foundation model SAM, known for its strong generalization ability, precisely segments cells within the countable regions of the BMA. The MAE, pre-trained on a large unlabeled cell dataset, excels in generalized feature extraction, enabling accurate classification of cells for counting. Additionally, TextureUnet and TCNet, with their powerful texture feature extraction capabilities, effectively segment the body-tail junction areas from WSIs and classify suitable tiles for DCCs. The framework was trained and validated on 40 WSIs from Chongqing Cancer Hospital. To assess its generalization ability across different medical centers and diseases, correlation tests were conducted using 13 WSIs from Chongqing Cancer Hospital and 5 WSIs from Southwest Hospital.

Results: The framework demonstrated high accuracy across all stages: The IoU for region of interest (ROI) segmentation was 46.19%, and the accuracy for tile of interest (TOI) classification was 90.45%, the Recall75 for cell segmentation was 99.01%, and the accuracy for cell classification was 77.92%. Experimental results indicated that the automated framework had excellent cell classification and counts performance, suitable for BMADCC across different medical centers and diseases. The differential cell counts results from all centers were highly consistent with manual analysis.

Conclusion: The proposed VFM-SSL-BMADCC-Framework effectively automates differential cell counts on bone marrow aspirate smears, reducing reliance on extensive annotations and improving generalization across medical centers.