Hemato-oncology

Challenge: High complexity and scarce data

Bone marrow smears contain a plethora of different cell types and maturity states
There are many different hemato-oncological diseases with finely branches sub-types
Bone marrow must be assessed in the context of a patient's comprehensive data

Our Approach: Integration of diverse data sources

Automation of differential cell counts in bone marrow
Make use of diverse data sets through foundation models
Integration of differential cell counts, clinical and lab data for fine-grained prediction of diseases
Facilitate the efficient creation of high-quality data sets

Research Topics

AI image analysis pipeline for detection and classification of bone marrow smear cells to generate differential cell counts — © Fraunhofer MEVIS
Image Analysis Pipeline for AI-based generation of differential cell counts in bone marrow aspirate smears.

Automation of differential cell counts

Differential cell counts of bone marrow smears are automatically created using AI in three steps
- Finding analyzable (monolayer) regions of a whole smear
- Cells within these regions are detected
- Detected cells are classified into a fine-grained set of cell types

References

[1] Selecting A Digitization Device for Bone Marrow Aspirate Smears. Kock et al. 2025. article

[2] Evaluating deep learning models for cell detection and multi-class cell classification: A comparative analysis of metrics and solutions. 70. Biometric Colloquium, Lübeck. Pfrang et al. 2024

[3] Transfer learning for cell detection in bone marrow smears. In Proceedings of the European Congress of Digital Pathology (ECDP). Kock et al. 2023.

[4] Comprehensive bone marrow analysis integrating deep learning-based pattern discovery (BMDeep). Pontones et al. 2022. article

Web-based cell labeling tool interface showcasing efficient dataset creation and high-quality annotation of cells — © Fraunhofer MEVIS
Creating high-quality datasets with our web-based labeling tool for efficient cell labeling.

Efficient creation of bone marrow cell data sets

Creation of data sets using an intuitive web application optimized for high annotation throughput
Rules can be set which user is shown which cell, e.g., to enable
- Minimum number of annotations per cell
- Minimum level of agreement among users
- Measurement of intra-observer agreement (with wash-out period)

Reference

[1] A workflow for efficient creation of high-quality cell classification data sets. In Proceedings of the European- Congress of Digital Pathology (ECDP). Höfener et al. 2023.

Innovative AI-based processing pipeline designed for leukemia type prediction. This comprehensive end-to-end system utilizes deep learning for cell analysis, enabling accurate diagnosis from digitized bone marrow aspirate smears. — Example of an end-to-end pipeline for deep learning-based cell analysis and subsequent diagnosis on a digitized bone marrow aspirate smear from a patient with chronic myeloid leukemia. (https://doi.org/g8wfcn, CC BY-NC-ND)

Prediction of diseases

Integration of laboratory data, clinical data and differential cell counts into join AI model
Prediction of disease types and sub-types

Reference

[1] Models for the marrow: A comprehensive review of AI-based cell classification methods and malignancy detection in bone marrow aspirate smears. Ghete et al. 2023. article