Deep Learning-based Processing of Hematological Image-data (DELPHI)

• Applicants: Professor Dr. Edgar Jost; Professorin Dr.-Ing. Dorit Merhof
• Subject Area: Epidemiology and Medical Biometry/Statistics
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 527820737

Digital cytology, which is a powerful tool for hematological analyses, comprises the detection and quantification of blood cells. For the diagnosis of hematopoietic diseases, knowledge of the cell type distribution in bone marrow is necessary. So far, this analysis is performed manually, as an automated analysis of hematopoietic cells comprises several technical challenges: bone marrow microscopy images show large variability in terms of cell appearance and staining across different samples. Furthermore, the large number of cell types as well as cell-like artifacts increases the difficulty of making accurate predictions. Additional challenges arise from handling (weak) annotations, a lack of annotated training data and class imbalance of cell types. In this project, we tackle the aforementioned challenges in order to automatically process bone marrow microscopy samples for hematological analyses. To this end, we utilize an already existing dataset arising from the collaboration of the two applicants that will be further extended during the course of the project. This dataset contains high quality digitized samples with annotations from medical experts. In the beginning of the project, we will perform research and propose solutions for typical problems related to the data basis, such as reducing the influence of variabilities in images and accelerating the annotation process. Afterwards, current limitation of state-of-the-art methods for classification and localization of hematopoietic cells will be addressed, e.g. by introducing new concepts for generative data augmentation. Another important aspect in hematological analyses is a measure to assess the confidence of predictions. Therefore, we will analyze techniques for computing such a confidence measure and also look into methods to incorporate prediction feedback from medical experts into the network training. To this end and also in order to include biological priors, hierarchical network structures and training processes will play a key role. Furthermore, we will take advantage of non-annotated data, which can be highly beneficial for network training due to their large number and variability. Finally, we will conduct inter- and intra-rater experiments for clinical validation as well as evaluations on non-hematological datasets to also assess how well our proposed approaches generalize to other kinds of data. In addition to publishing papers and journals, we plan to make the source code of our proposed methods as well as the dataset including annotations available to the scientific community. Our project DELPHI will therefore not only provide automated solutions for digital cytology of bone marrow samples, but will also provide insight into common problems encountered in medical image analysis.

• DFG Programme: Research Grants