Cross-Domain Few-Shot Sparse-Quantization Aware Learning for Lymphoblast Detection in Blood Smear Images

Deep learning for medical image classification has enjoyed increased attention. However, a bottleneck that prevents it from widespread adoption is its dependency on very large, annotated datasets, a condition that cannot always be satisfied. Few-shot learning in the medical domain is still in its infancy but has the potential to overcome these challenges. Compression is a way for models to be deployed on resource-constrained machines. In an attempt to tackle some of the challenges imposed by limited data and high computational resources, we present a few-shot sparse-quantization aware meta-training framework (FS-SQAM). The proposed framework aims to exploit the role of sparsity and quantization for improved adaptability in a low-resource cross-domain setting for the classification of acute lymphocytic leukemia (ALL) in blood cell images. Combining these strategies enables us to approach two of the most common problems that encounter deep learning for medical images: the need for extremely large datasets and high computational resources. Extensive experiments have been conducted to evaluate the performance of the proposed framework on the ALL-IDB2 dataset in a cross-domain few-shot setting. Performance gains in terms of accuracy and compression have been demonstrated, thus serving to realize the suitability of meta-learning on resource-constrained devices. Future advancements in the domain of efficient deep learning computer-aided diagnosis systems will facilitate their adoption in clinical medicine.