Automatic identification of blast leukemic cells types using deep learning with vision transformers.

Abstract Background: Acute myeloid leukemia (AML) and acute lymphoblastic leukemia (ALL)are heterogeneous hematopoietic neoplasms with poor outcomes, their diagnosis require holistic integration of clinical history, morphology, immunophenotype and cytogenetic/molecular genetic analysis.The diagnosis of AML and ALL in Colombia, a middle-income country has significant challenges due to delays in diagnosis confirmation, which can range from 9 to 24 days, or even longer in areas with limited access to healthcare services with negative impacts in the patient's outcomes. Images analysis throw deep learning using vision transformers techniques can be affordable strategies for developing diagnosis support models for acute leukemias. This study aims to present results of an Artificial intelligence (AI)-based model to differentiate myeloblasts from lymphoblast. Methods: This was a retrospective, diagnosis study conducted in the Hospital de San Jose in Bogotá analyzing bone marrow smears of 24 acute myeloid leukemia, 19 acute lymphoblastic leukemia and 5 acute promyelocytic leukemia (APL) cases attended between 2021 and march 2025. Clinical information of adult patients (≥18 years) diagnosed with AML, ALL and APL was analyzed including demographics, disease diagnosis, treatment and outcome data. The bone marrow smears of the cases were scanned with high resolution scanner Ventana DP200, then those samples were manually annotated in the QuPath software and curated by two expert morphologists. Then images were processed to enhance visual characteristics like high definition preserving their distinctive features. A Phyton-based code was developed for the recognition of blast cells, emphasizing in differentiation between myeloblasts and lymphoblasts using PyTorch library. A second task, the classification of acute promyelocytic leukemia (APL) cases from other AML cases was also analyzed. Five models were trained, two of them based on vision transformers (ConvNeXt Small/Large and Swin Transformer Tiny) and the other three models based on convolutional neural networks (EfficientNet-B0, VGG16 and ResNet50). Evaluation was performed with the validation set. Descriptive statistics, accuracy metrics and area under the curve (AUC-ROC) are reported. Results: The median age at diagnosis was 53±16 years (range, 26-80) in AML, 40±18 (range, 20-76) years for ALL, and 53±16 years for APL, with a predominance of males (75% in AML cases, 57% in ALL cases and 80% en APL cases). At diagnosis, hyperleukocytosis was present in 42% of ALL cases. A total of 5454 images from AML and 4689 from ALL were used for training; 606 images from AML and 522 from ALL for validation. The best accuracy for morphological differentiation between myeloblasts and lymphoblast was performed by the models based on vision transformers, the ConvNeXt small model showed 99.67% sensitivity, 98.58% specificity, 99.02% value predictive positive and 99.61% negative predictive value with an AUC-ROC of 0.998. The models EfficientNet B0 and ConvNeXt Small were the least accurate for the differentiation of APL whereas models like ConvNeXt Large, which has a higher-capacity architecture achieved a 64.58% accuracy and an AUC-ROC of 0.9407. Conclusions: The results obtained in this study show that the AI-based bone marrow morphological analysis system can be a tool for the successful differentiation of myeloblasts and lymphoblasts, specially implementing vision transformers-based models. Large-scale validation studies are warranted to further validate the clinical utility of this system.