Accurate subtyping of acute leukemia is essential for guiding therapy and predicting patient outcomes. Morphological assessment remains challenging for distinguishing subtypes with subtle cytomorphologic differences, particularly in rare or atypical forms where reliable classification is limited. Recent computational models have attempted to automate this process. However, their clinical applicability was limited by insufficient generalizability and granularity across subtypes of acute leukemia. Here we developed a deep learning framework for automated cell-level classification and case-level subtyping of acute leukemia from Wright-Giemsa-stained bone marrow smears. The model was trained on 180,928 expert-annotated single-cell images representing 19 hematopoietic and leukemic cell categories collected from three different imaging platforms to enhance generalizability. ALSNet incorporates a dual-branch convolutional architecture and a Transformer encoder to capture both fine-grained local features and global morphological context. Internally, ALSNet achieved per-class accuracies up to 0.99 for mature cells and > 0.80 for diagnostically relevant precursors, while in an external validation from an independent platform, case-level accuracy reached 0.75 with leukemic cell percentage strongly correlated to manual review (R2 = 0.66). These results indicate that ALSNet enables robust, platform-independent morphological classification and may facilitate the early, reliable diagnosis of acute leukemia in clinical practice.

Introduction: Accurate morphological classification of acute leukemia (AL) subtypes from bone marrow smears (BMSs) remains a cornerstone of the MICM (morphology–immunophenotyping–cytogenetics–molecular biology) diagnostic framework and is critical for guiding clinical decisions and prognostic assessment. However, manual evaluation of leukemic cell morphology is labor-intensive, time-consuming, and highly dependent on expert hematopathological interpretation. Although recent artificial intelligence (AI)-based methods have shown promise in automating cell recognition, existing approaches are often limited by their narrow coverage of leukemic subtypes and insufficient generalizability across different imaging platforms. To overcome these limitations, we developed a large diagnosis-oriented single cell image dataset and trained ALSNet (acute leukemia subtyping network), a cross-platform deep learning model designed for automated and robust subtyping of AL using region-of-interest images derived from BMSs. Methods: A total of 1,232 Wright–Giemsa–stained bone marrow smears (BMSs) from patients at initial diagnosis were collected and digitized under 100× oil immersion using both manual microscopes and two automated scanning platforms. The dataset included 210 normal controls, 210 acute lymphoblastic leukemia (ALL), 150 acute promyelocytic leukemia (APL), and 662 acute myeloid leukemia (AML) cases. Cell annotation was performed using a human-in-the-loop framework, integrating model-assisted pre-classification to improve efficiency and consistency. A Mask R-CNN–based detection and segmentation network was first employed to extract single-cell images from raw multi-cell fields. A CNN based on the GoogLeNet architecture was trained to exclude low-quality images containing artifacts, ruptured cells, or debris. To ensure generalizability across platforms, CIELAB-based histogram matching was used for color normalization. After preprocessing, single-cell images were categorized into 19 cytomorphologic classes relevant to leukemia diagnosis. These images were used to train ALSNet, a ResNeXt-based deep learning model with a dual-branch Transformer module, to perform cell-level classification and leukemia subtyping according to the WHO classification. Subtyping encompassed ALL, APL, and diverse AML subtypes, including AML without maturation, AML with maturation, acute myelomonocytic leukemia (AMML), acute monoblastic leukemia (AMBL), acute monocytic leukemia (AMOL), acute erythroid leukemia (AEL), and acute megakaryoblastic leukemia (AMKL). Model performance was validated on an independent testing cohort consisting of 100 AL cases and 10 normal controls imaged using a separate scanning platform. Results: Analysis of clinical characteristics revealed different baseline features among leukemic subtypes, such as more significant thrombocytopenia in APL and AMKL. Segmentation and quality control (QC) modules showed high accuracy (AUC = 0.9772 for segmentation; AUC = 0.9926 for QC filtering). Following expert-verified annotation, a total of 182,230 high-quality single-cell images were compiled to establish a large-scale cytomorphologic training dataset. The ALSNet model achieved a macro-averaged accuracy of 0.912 in cell-level classification across 19 categories. At the case level, the model correctly identified normal samples, ALL, APL and AML subtypes with an overall accuracy of 0.75 in the testing set, demonstrating cross-platform robustness.Conclusions: This study presents ALSNet, a cross-platform AI model capable of automated, high-accuracy subtype classification of AL from routine BMS images. By integrating robust image preprocessing, single-cell recognition, and advanced classification architecture, ALSNet addresses key limitations in previous AI diagnostic efforts. The model's generalizability across imaging systems and its alignment with clinical phenotypes support its potential for real-world deployment in hematologic diagnostics.

Abstract Introduction: Morphologic evaluation of bone marrow cells remains a cornerstone of the MICM (morphology-immunophenotyping-cytogenetics-molecular biology) diagnostic framework in hematologic malignancies. In plasma cell neoplasms, including multiple myeloma (MM) and plasma cell leukemia (PCL), neoplastic plasma cells exhibit marked morphological heterogeneity. Characteristically, plasma cells are defined by abundant basophilic cytoplasm and eccentrically positioned nuclei, reflecting their role as terminally differentiated, immunoglobulin-secreting cells. However, PCL, as a highly aggressive form of plasma cell neoplasm, often displays distinct morphological features compared to typical MM. To date, the potential association between plasma cell morphology and clinical prognosis has not been systematically investigated. This study explored the potential prognostic value of plasma cell morphology in patients with plasma cell dyscrasias. Methods: This retrospective study included 347 newly diagnosed multiple myeloma (MM) patients and 30 patients with either primary or secondary plasma cell leukemia (PCL). Wright-Giemsa-stained bone marrow smears (BMSs) and corresponding clinical data were collected at diagnosis. All BMSs were scanned under 100× oil immersion using an automated imaging platform, with a minimum of 500 nucleated cells captured per case. Following standardized pipelines for cell detection, segmentation, and quality control, plasma cells from MM and PCL cohorts were automatically identified using a previously trained cell classification model (unpublished). Pixel-level segmentation masks for nuclei and whole-cell boundaries were generated using a U-Net–based deep learning algorithm, enabling quantitative measurement of nucleus-to-plasma ratio and nuclear skewness for each plasma cell. The mean values of these morphologic parameters were calculated per case and analyzed in relation to clinical baseline characteristics for collinearity. Correlations with continuous variables were assessed using Spearman's rank correlation, while non-parametric tests were applied for categorical comparisons. Finally, morphology-related variables and clinical indicators associated with prognosis were evaluated using multivariate Cox regression analysis. Results: The plasma cell classification model achieved a high accuracy of 0.9747 with an AUC of 0.9965. The segmentation model for nucleus and cytoplasm yielded a precision–recall AUC of 0.9721, with bounding box errors predominantly within 5%, ensuring reliable morphologic quantification. The accuracy of both the classification and subcellular segmentation models provided a robust foundation for the downstream extraction of morphological features. Comparative analysis of plasma cell morphology between MM and PCL patients revealed significantly elevated nucleus-to-plasma (N:C) ratios and decreased nuclear eccentricity in the PCL cohort. Mann-Whitney U tests further identified sex-associated differences in morphology among MM patients, with male patients exhibiting higher N:C ratios and lower eccentricity compared to females. Moreover, both N:C ratio and nuclear eccentricity were negatively correlated with hemoglobin levels, indicating that plasma cells with higher N:C ratios and lower eccentricity were associated with more severe anemia. Multivariate Cox regression analysis based on available survival data demonstrated that a higher N:C ratio was independently associated with inferior progression-free survival (HR = 2.12; 95% CI: 1.38-3.29; p = 0.042), while increased nuclear eccentricity served as a protective factor (HR = 0.73; 95% CI: 0.38–0.85; p = 0.035). The overall model was statistically significant with a concordance index (C-index) of 0.72.Conclusions: This study demonstrates that quantification of plasma cell morphology from bone marrow smears enables precise measurement of nucleus-to-plasma ratio and nuclear eccentricity. Notably, higher nucleus-to-plasma ratio and lower nuclear eccentricity were associated with more aggressive disease features and inferior hematologic profiles. Furthermore, these morphologic parameters independently predicted patient survival in MM, highlighting their potential as novel prognostic biomarkers. Integration of AI-assisted morphologic assessment into routine diagnostics may enhance risk stratification and guide therapeutic decision-making in clinical practice.

Most existing deep learning-based segmentation approaches for color medical images tend to process color and texture data together, overlooking the distinct distribution patterns of color information, such as channel-wise variations and spectral correlations. To address this limitation, we present CoSegNet, a novel network specifically designed for color medical image segmentation. CoSegNet incorporates the mathematical framework of quaternion to develop two key components: the Quaternion Bottleneck Module (QBottle) and the Quaternion Hierarchical Self-Attention Mechanism (QHSA). The QBottle module is a quaternion-based local feature extraction component designed to reduce computational complexity by compressing feature dimensions while preserving critical color information. On the other hand, the QHSA module is specifically tailored to capture global color features with long-range dependencies. Unlike the traditional self-attention mechanism, QHSA captures the long-range dependencies of the target in the color space by deflating and interacting with quaternion-based color representations in different dimensions. Furthermore, we propose a Quaternion Cross-Semantic Restrainer (QCSR), which can align multi-level color features in the color space, addressing semantic discrepancies during the fusion of these features. Experimental results show that CoSegNet surpasses state-of-the-art methods on four medical image datasets, achieving mDice scores of 91.24 % (GlaS), 85.24 % (CVCClinicDB), 90.27 % (ISIC 2018), and 92.43 % (2018 DSB). Notably, CoSegNet has approximately 75 % fewer parameters than traditional U-Net architectures. Our code will be available at https://github.com/chasingone/CosegNet.

<p>PDF file - 179KB, Sequence of oligonucleotides used in this study.</p>

Supplementary Figure Legends. Legend for Supplementary Figures S1-S4.

<p>PDF file - 2673KB, MiR-30s target BCL9 in MM cells.</p>

PDF file - 141KB

PDF file - 2071KB, MiR-30s treatment does not have a negative impact on bone disease in MM1S-Luc-Neo bearing mice.

Supplementary Figure S3. Changes in PUMA mRNA expression induced by V-miR-221-222-GFP and V-as-miR-221-222-GFP.

<p>PDF file - 2673KB, MiR-30s target BCL9 in MM cells.</p>

Supplementary Figure S1. Western blot evidence of GR and PUMA downregulation in MM1R cells as compared with MM1S cells.

<p>PDF file - 1007KB, Knockdown expression of miR-30s enhances BCL9 expression in MM cells.</p>

<p>PDF file - 1392KB, BMSCs decrease miR-30s levels, and enforced expression of miR-30s inhibits CAM-DR in MM cells.</p>

<p>PDF file - 1007KB, Knockdown expression of miR-30s enhances BCL9 expression in MM cells.</p>

Supplementary Figure S2. AGO2/RNA immunoprecipitation assay in MM1S cells.

Supplementary Figure S4. Increased drug resistance to apoptosis in V-miR-221-222-GFP stable transduced MM1S cells.

Supplementary Table S1. MM patients' information.

PDF file - 179KB, Sequence of oligonucleotides used in this study.

Subsequently to the publication of the above article and a Corrigendum that was published to indicate corrections made to Fig. 7 (DOI: 10.3892/or.2021.7922; published online on January 5, 2021), a concerned reader drew the Editor's attention to the fact that, comparing between a pair of panels in the Figure, there was an overlapping section of data; moreover, this overlapping section contained apparent anomalies that could not be easily accounted for through a straightforward re‑use of one of the data panels. The authors conceded that there was partial duplication between the images shown in Fig. 7B and F, although they were unable to access the related raw data as the experiments had been performed almost 10 years ago. Secondly, the authors informed the Editor that the corresponding author did not know he was on the author list at the time of submission. Although the authors' were granted permission to publish the Corrigendum, the Editor now considers that the paper should be retracted on account of the uncertainties in the presented revised data and the authors' admission concerning the corresponding author. Therefore, this paper has been retracted from the Journal. The authors are in agreement with the decision to retract the article. The Editor apologizes for any inconvenience caused. [Oncology Reports 37: 2751‑2760, 2017; DOI: 10.3892/or.2017.5569].