Automated classification of acute lymphoblastic leukemia in peripheral blood smears using a convoluted neural network, ResNet18 model

Abstract

Abstract Background Acute lymphoblastic leukemia (ALL) is the most prevalent malignancy in children and remains a therapeutic challenge in adult populations. While genomic profiling and molecular subtyping have propelled precision risk stratification, the morphological evaluation of leukemia remains heavily reliant on expert interpretation, introducing variability and time constraints. Recent advances in deep learning, particularly convolutional neural networks, offer powerful modalities to automate cytomorphological analysis, reducing subjectivity and accelerating diagnostic workflows. We describe the development and validation of a ResNet18-based algorithm embedded within a cross-platform application designed to classify ALL subtypes with near-perfect accuracy and sub-second inference times. Methodology A convolutional neural network based on the ResNet18 architecture was trained using an anonymized dataset of hematoxylin and eosin-stained peripheral blood smear images, sourced from a publicly available repository. The dataset was partitioned into training (60%), validation (20%), and testing (20%) cohorts according to standardized protocols. Data augmentation strategies were applied to mitigate overfitting and enhance generalizability. The network was optimized using stochastic gradient descent with momentum, employing a learning rate scheduler to refine convergence. After internal validation, the final model was deployed within a secure, cross-platform interface and independently assessed by clinicians from international institutions to confirm usability and real-world performance. Results On the held-out test cohort, the ResNet18 model attained an overall classification accuracy of 98.8% (483/489). Breakdown by category revealed accuracies of 100.0% for early-stage pre-B ALL and pre-B ALL blasts, 98.4% for pro-B ALL blasts, and 91.8% for benign lymphocytes. The confusion matrix highlighted three instances of benign lymphocytes misclassified as early-stage pre-B blasts and one mislabeled as a pro-B blast; two pro-B blast images were erroneously assigned to the pre-B category. Receiver operating characteristic curves yielded areas under the curve of 1.00 for all four cell classes, and precision-recall analyses demonstrated precision scores of 1.00 across each category. Training and validation accuracy converged toward unity by epoch 5, while loss approached zero. The application processed each image in under one second, underscoring its potential for real-time clinical integration. Discussion The ResNet18-driven classification application exhibits exceptional discriminative capability across ALL subtypes, mirroring human expert performance while delivering exceptional consistency. The slight decrement in benign lymphocyte specificity underscores the necessity for ongoing refinement and integration with cytogenetic or molecular assays to ensure diagnostic robustness. Nonetheless, the model's rapid inference and high accuracy position it as a transformative tool for augmenting hematopathology workflows. Its deployment could democratize access to expert-level leukemia classification in under-resourced settings, streamline tele-hematology consultations, and standardize reporting across institutions. Future work will focus on integrating multimodal data streams, such as flow cytometry and genomic profiles, to further enhance subtype resolution and prognostic stratification.