Abstract Background: Acute lymphoblastic leukemia (ALL) represents a clonal expansion of immature lymphoid progenitors and remains the predominant pediatric hematologic malignancy, while posing significant therapeutic and diagnostic hurdles in adults. Contemporary workflows increasingly integrate genomic profiling for risk stratification, yet cytomorphological inspection of peripheral blood and bone marrow smears endures as the de facto initial diagnostic modality, albeit one that is inherently subjective and labor-intensive. Advances in deep convolutional neural networks promise to automate and standardize morphologic evaluation, thereby expediting diagnosis and reducing inter-observer variability. Objective: To architect and validate an EfficientNetB1-powered AI application for accurate diagnosis and prognosis of acute lymphoblastic leukemia, with particular emphasis on computational efficiency, data privacy, and equitable deployment in resource-constrained environments. We aim for high-fidelity discrimination of lymphocytes from the various ALL subtypes (early-stage pre-B, pre-B, pro-B). Methods: A curated, de-identified dataset of hematoxylin and eosin-stained peripheral blood smears from confirmed ALL cases and healthy controls was partitioned into 60% training, 20% validation, and 20% test cohorts. The EfficientNetB1 model was trained using transfer learning, fine-tuned with stratified class weighting, and regularized via early stopping. Post-training, the model was incorporated in a secure, user-centric application tested for usability and reliability by clinicians from several healthcare institutions across the globe. Results: On an independent test set of 2,436 images, the EfficientNetB1–powered AI application achieved perfect classification across all four categories, yielding an overall accuracy of 100% (2,436/2,436). Class-specific receiver operating characteristic curves each attained an area under the curve of 1.00, and precision-recall analyses returned average precision of 1.00 per class. Training and validation accuracy curves converged to unity within the first ten epochs without notable overfitting, as corroborated by near-zero loss values. The confusion matrix demonstrated exclusively diagonal entries, confirming zero misclassifications. Inference time per image remained under one second, affirming the platform's suitability for real-time screening. Conclusions: The AI application leveraging EfficientNetB1 furnishes flawless discrimination of ALL subtypes, coupling unparalleled diagnostic precision (100% accuracy, AUROC=1.00) with sub-second inference. These attributes position this application as a transformative tool for point-of-care cytomorphological assessment, facilitating rapid, objective, and reproducible ALL diagnosis. Discussion: These findings substantiate the feasibility of deploying deep learning–driven cytomorphology pipelines to mitigate observer bias and reduce diagnostic timelines. Continued refinement, incorporating diverse staining protocols, multicenter data augmentation, and prospective clinical validation will be pivotal to ensuring longitudinal robustness and seamless integration within laboratory information systems. Ultimately, this AI application promises to democratize access to high-precision leukemia diagnostics, particularly in underresourced settings, thereby advancing global equity in hematological care.
Elangovan et al. (Mon,) studied this question.