Peripheral blood smear analysis: A comparison of GPT‐4o and supervised deep learning models

Peripheral blood smear review remains pivotal to hematologic diagnosis.Morphologic abnormalities central to contemporary hematologic classification, including blasts, dysplastic neutrophils, nucleated red blood cells, and abnormal platelets, remain critical to diagnosis under both the World Health Organisation 1 and the International Consensus Classification frameworks. 2 While automated analysers have improved workflow and standardisation in cell counting, they often fall short in detecting key morphological features.These diagnostic gaps make smear interpretation by trained personnel indispensable in both routine and urgent haematology.However, access to expert haematological review is variable, and manual interpretation is time-intensive and subject to interobserver variability.This has led to increasing interest in artificial intelligence (AI) systems that can support or automate morphologic assessment by learning from large sets of labelled smear images.Recent advances in machine learning have enabled the development of models capable of classifying peripheral blood cells from digital images with high accuracy.Convolutional neural networks and transformer-based architectures have shown promise in this domain, achieving expert-level performance when trained on large, curated datasets.3 However, these supervised models rely on thousands of annotated examples for each cell type and often require fine-tuning before deployment.In most clinical settings, such resources are not readily available, and differences in staining, imaging, and labelling practices further limit generalizability across institutions.As a result, the practical application of supervised AI models remains dependent on the availability of large amounts of annotated data in addition to high-performance computing.Therefore, to address these limitations, recent studies have explored models that require far less annotated data to perform diagnostic tasks.One such approach is fewshot learning, where the model is shown a small number of labelled examples at the time of evaluation rather than