Loading...
Novel hierarchical deep learning models predict type of leukemia from whole slide microscopic images of peripheral blood
Syed, Naveed Saeed, Mohamed Eltag Salih Hussain, Shakir Mirza, Imran Abdalla, Amira Mahmoud Al Zaabi, Eiman Ahmed Afrooz, Imrana Hashmi, Shahrukh Yaqub, Mohammad
Syed, Naveed
Saeed, Mohamed Eltag Salih
Hussain, Shakir
Mirza, Imran
Abdalla, Amira Mahmoud
Al Zaabi, Eiman Ahmed
Afrooz, Imrana
Hashmi, Shahrukh
Yaqub, Mohammad
Files
Loading...
jmai-08-5.pdf
Adobe PDF, 1.64 MB
Supervisor
Department
Computer Vision
Embargo End Date
Type
Journal article
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Background: Diagnosis of leukemia relies on microscopic examination of suspected peripheral blood smears (PBS) by pathologists. However, this challenging and potentially error-prone process is further limited by the availability of pathologists. Developing a clinically applicable predictive artificial intelligence (AI) model offers a potential solution to these limitations. This study aimed to develop an AI model that predicts leukemia from microscopic images of peripheral blood smear.
Methods: We developed hierarchical deep learning (DL) pipelines that were trained and validated on 7,255 high-power field (HPF) microscopy images collected from pre-diagnosed cases at our centre between January 2021 to August 2023. The proposed pipelines employed a high performing binary DL model in step 1 and multiclass DL models in step 2 to predict the label of each image from a slide, and then aggregate the results to provide a patient-level diagnostic prediction. This approach mimics the examination of PBS by trained hematopathologists, who examine multiple HPF patches under a microscope to predict the type of leukemia. The pipelines’ performance was assessed using F1 score, sensitivity, specificity, and accuracy, then compared against the three pathologists using Cohen’s kappa score.
Results: The 3-class pipeline categorizes the slide as acute leukemia, chronic leukemia, or no leukemia. The F1 score, sensitivity, specificity, and accuracy were 94%, 94%, 95%, and 92% respectively. This pipeline outperformed the pathologists in all metrics. The agreement between this pipeline and pathologists ranged between 74–77%, while the agreement amongst the pathologists ranged from 82–87%. The 5-class pipeline categorizes the slide into acute myeloid leukemia (AML), acute lymphoblastic leukemia (ALL), chronic myeloid leukemia (CML), chronic lymphoblastic leukemia (CLL), and no leukemia. It achieved an F1 score of 77%, sensitivity of 80%, specificity of 94%, and accuracy of 70%. The 5-class pipeline outperformed pathologists in the CML and CLL classes but was inferior in other classes. The agreement between this pipeline and pathologists ranged between 52–53%, while the agreement amongst the pathologists ranged from 72–79%.
Conclusions: Acute and chronic leukemia can be predicted with high sensitivity and specificity from whole-slide microscopic images of PBS by utilizing three deep-learning models in a hierarchical pattern and aggregating the results. This approach can efficiently predict and differentiate CLL from CML, but not
ALL from AML. Applying cell segmentation in this level of the pipeline could improve the prediction in these classes. These pipelines can be applied in limited infrastructure settings to potentially identify types of leukemia with greater reliability.
Citation
N. Syed et al., “Novel hierarchical deep learning models predict type of leukemia from whole slide microscopic images of peripheral blood,” J Med Artif Intell, vol. 8, no. 0, Mar. 2025, doi: 10.21037/JMAI-24-74/COIF.
Source
Journal of Medical Artificial Intelligence
Conference
Keywords
Acute myeloid leukemia (AML), acute lymphoblastic leukemia (ALL), chronic myeloid leukemia (CML), chronic lymphocytic leukemia (CLL), deep learning (DL)
Subjects
Source
Publisher
AME Publishing Company