FEW-SHOT LEARNING FOR AML CELL CLASSIFICATION USING PROTOTYPICAL NETWORKS

Accurate blood cell classification is crucial for diagnosing Acute Myeloid Leukemia (AML) but limited medical data poses challenges for traditional machine learning models. This study presents a Few-Shot Learning (FSL) framework utilizing a Prototypical Network architecture with a ResNet-34 backbone to classify AML blood cell types from microscopic images. In this study, we utilize datasets consisting of 15 morphologically distinct cell classes. A 15-way, 5-shot, 5-query episodic setup was adopted to simulate data-scarce conditions. Evaluation via 5-fold cross-validation yielded strong performance, with an average accuracy of 97.76%, precision of 98.78%, recall of 96.55%, and F1-score of 97.76%. FSL training times were consistent (4.22–4.26 minutes per fold), and t-SNE along with confusion matrices confirmed the model’s ability to distinguish similar cell types. To validate the approach, its performance was compared with a conventional supervised CNN using the same ResNet-34 backbone. The FSL model outperformed the CNN across all metrics such as accuracy (98.32% vs. 77.25%), precision (98.55% vs. 76.87%), recall (98.31% vs. 78.66%), and F1-score (98.33% vs. 75.26%), while also requiring far less training time (~4.24 min/fold vs. ~420 min total). These results highlight the promise of FSL based methods for accurate, efficient, and scalable hematologic diagnostics in data limited settings.