AI-powered leukemia intelligent virtual examiner (LIVE) for rapid and accurate diagnosis of acute lymphoblastic leukemia

Abstract Background: Acute lymphoblastic leukemia (ALL) is a hematological malignancy characterized by clonal proliferation of immature lymphoid progenitors. ALL is the most common malignancy in pediatric populations and presents substantial therapeutic challenges to adult populations. Contemporary diagnostic methods heavily rely on sophisticated and expensive molecular subtyping and genomic profiling for precise risk stratification whereas traditional morphological assessment remains subjective and labor-intensive. This makes it harder to diagnose ALL in resource-limited settings. Artificial intelligence is revolutionizing healthcare diagnostics through automated pattern recognition. Particularly, deep learning algorithms hold great promise in hematological malignancies. State-of-the-art convolutional neural networks have demonstrated remarkable accuracy in ALL detection and classification. These AI-driven innovations herald transformative opportunities for standardized, objective, and rapid ALL diagnosis across diverse clinical settings. Aim: To develop and validate a computationally efficient, AI-powered crossplatform application, Leukemia Intelligent Virtual Examiner (LIVE), utilizing the EfficientNetB0 architecture for accurate detection of acute lymphoblastic leukemia, with a focus on privacy, real-world usability, and equitable access in underserved settings. Methodology: We developed and trained an AI model based on the EfficientNetB0 architecture using an anonymized dataset of hematoxylin and eosin-stained peripheral blood smears from leukemic patients. The dataset was partitioned into training (60%), validation (20%), and testing (20%) sets. The model was trained, internally validated and tested using the apportioned test set. A secure, user-friendly crossplatform application was created for morphometric analysis. Its reliability and usability were independently validated by healthcare professionals across the globe, ensuring real-world applicability and global generalizability. Results: The EfficientNetB0 model achieved a per-class accuracy of 100% for pre-B and pro-B ALL and 99.3% for early-stage pre-B, with an overall accuracy of 99.8% (488/489) on the apportioned test set. The confusion matrix indicated 76 true benign, 146 true early-stage pre-B, 150 true pre-B, 116 true pro-B classifications, and a single misclassification of an early-stage pre-B smear as benign. Receiver operating characteristic curves yielded class-specific AUROCs of 1.00, and precision-recall analysis demonstrated average precision scores of 1.00 across all categories. Evaluation in our crossplatform application required less than a second per image, confirming rapid and robust performance. Conclusions: The EfficientNetB0-powered application delivers near-perfect classification of ALL subtypes, achieving 99.8% overall accuracy and AUROCs of 1.00. Its sub-second inference and flawless precision-recall performance demonstrate both diagnostic accuracy and computational efficiency. Discussion: These results underscore the clinical viability of AI-driven cytomorphological analysis, mitigating observer variability and accelerating diagnostic workflows. The single misclassification observed highlights the importance of continuous model refinement and integration with laboratory quality control processes. Applications: The AI-powered Leukemia Intelligent Virtual Examiner enables point-of-care ALL screening in resource-limited settings, preserving patient privacy and facilitating equitable access. Its adoption can standardize diagnoses, expedite treatment decisions, and support telehematology across global healthcare networks.