BloodImage: Benchmarking vision transformers for blast detection in digital blood films using public and clinical datasets.

Background and objectives: Leukemia is one of the most common cancers in the UK and it is usually initially diagnosed through the time-consuming and subjective analysis of blood films by an expert hematologist. When a small number of blast cells may be present on a blood film, it is difficult to detect them even after a thorough review. Automating blood film image analysis could significantly speed up the process and improve diagnostic accuracy. This study benchmarks a machine learning framework based on vision transformers (ViTs) for automated blast detection in digitized blood films, evaluating their generalizability across public and clinical datasets. Methods: We investigated different training strategies (hold-out/k-fold cross-validation), optimization (Adam or stochastic gradient descent (SGD)), and data preprocessing techniques (data augmentation, Gaussian pyramid downsampling) to assess their impact on the ViT performance when tested using both public (ALL-IDB) and clinical datasets from Barts Health NHS Trust. Results: Models trained with Adam performed better than those trained with SGD. The best-performing model, ViT2-Adam, achieved the highest accuracy (≥0.86) and area under the receiver operating characteristic curvearea under the curve (AUROC ≥ 0.95), which exceeded other stochastic models demonstrating its potential for integration into clinical diagnostic workflows. Conclusions: Our findings support the viability of ViTs for clinical integration in blood film analysis. Augmentation, advanced data splitting, and Gaussian downsampling enhance model generalization, offering a promising strategy for resource-limited or high-throughput diagnostic environments.