Home Publications Research Grants Inventions & Patents Awards Additional Research Activities Faculties & Institutions Research Areas

UNSTAINED WHITE BLOOD CELL CLASSIFICATION USING DEEP LEARNING

Yu H, Forster F, Bhandary Panambur A, Merino A, Maier A, Marquardt G (2023)

---

Publication Language: English

Publication Type: Conference contribution

Publication year: 2023

Journal

Publisher: Wiley

Edited Volumes: Abstracts: UNSTAINED WHITE BLOOD CELL CLASSIFICATION USING DEEP LEARNING

Series: International Journal of Laboratory Hematology

Book Volume: 45 S3

Pages Range: 3-137

Conference Proceedings Title: Abstracts of International Society for Laboratory Hematology XXXVIth International Symposium on Technical Innovations in Laboratory Hematology

Event location: New Orleans, LA

Journal Issue: 45

URI: https://onlinelibrary.wiley.com/doi/10.1111/ijlh.14149

DOI: 10.1111/ijlh.14149

Open Access Link: https://onlinelibrary.wiley.com/doi/epdf/10.1111/ijlh.14149

Abstract

Introduction: Hematological diseases, including leukemia, lymphoma, etc., are associated with White Blood Cell (WBC) disorders. Therefore, precise classification of WBCs is critical in peripheral blood (PB) smear analysis. However, the traditional staining process on the smears is time-consuming and expensive. Additionally, the chemical residuals are detrimental to the operators and the environment. In this work, we develop a two-stage deep learning (DL) pipeline for the classification of 14 different types of unstained WBC. Methods: PB smears of 233 different patients were collected from two different hospitals. These PB smears were microscopically digitalized before and after staining and a dataset consisting of both normal and abnormal WBC digital images was acquired. The stained cell images were initially annotated by two clinical experts. These annotations were then transferred to the unstained ones by alignment. A 14 different classes of blood cells have been annotated. The 14 types of WBC images are as follows: Basophil granulocyte (BA), blast (BL), eosinophil granulocyte (EO), immature granulocyte (IG), neutrophil granulocyte (NE), lymphocyte (LY), abnormal lymphoid cell (ALC), monocyte (MO), reactive lymphocyte (RL), plasma cell (PC), nucleated red blood cell (NRBC), giant platelet (GPLT), smudge (SMU) and artifact (ART). The total number of unstained images is 339 885. A 70%, 17%, and 13% of the total dataset were used for training, validation, and testing, respectively. As shown in Fig. 1, a patient-wise classification strategy with a transformer was designed to improve the classification rate using a pre-trained ResNet-18 as the feature extractor. This was performed to potentially mimic how pathologists classify by comparing different cells of the same patient to classify the normal and pathological cells. Results: Performing the evaluation on the testing set for both ResNet-18 and transformer models, mean F1 weighted scores of 90.33% and 91.77% are achieved, respectively. Thus, transformer approach with patient-wise method outperforms ResNet-18 by 1.44%. Fig. 2 shows the density confusion matrix of the best model with patient-wise transformer strategy. Even though the classification of normal WBCs is robust, the identification of some abnormal types still requires improvement. Conclusion: The proposed two-stage DL method shows promising results on unstained WBC classification. The transformer approach improves the classification performance on top of conventional ResNet-18 using patient-wise strategy.

Authors with CRIS profile

Involved external institutions

How to cite

APA:

Yu, H., Forster, F., Bhandary Panambur, A., Merino, A., Maier, A., & Marquardt, G. (2023). UNSTAINED WHITE BLOOD CELL CLASSIFICATION USING DEEP LEARNING. In Abstracts of International Society for Laboratory Hematology XXXVIth International Symposium on Technical Innovations in Laboratory Hematology (pp. 3-137). New Orleans, LA, US: Wiley.

MLA:

Yu, Hui, et al. "UNSTAINED WHITE BLOOD CELL CLASSIFICATION USING DEEP LEARNING." Proceedings of the XXXVIth International Symposium on Technical Innovations in Laboratory Hematology, New Orleans, LA Wiley, 2023. 3-137.

BibTeX: Download