Robustness Assessment of Feature Selection-Based Hybrid Deep Learning under Sample Size Transition from Small to Large

Abstract

Although deep learning exhibits high generalization performance, it requires large and diverse datasets. However, in practical implementations, rapid deployment in real-world settings is sometimes required even under constraints that limit data availability. In this context, deep learning models that maintain strong generalization with limited data and continue to improve as more data are collected are desirable. In this study, we focus on a feature selection-based hybrid deep learning approach and assess its robustness to the transition from small to large sample sizes in a binary classification task of detecting tumor tissue in lymph node histopathology images. The findings may enable early deployment of models for operational support and improve sustained performance.