Abstract

In deep-learning-based image classification, achieving high performance requires diverse training sets. However, the current best practice—maximizing dataset size and class balance—does not guarantee dataset diversity. We hypothesized that, for a given model architecture, performance improves by maximizing diversity more directly. To test this hypothesis, we introduce a comprehensive framework of “sentropic” diversity measures or “sentropies” from ecology that generalizes familiar quantities like Shannon entropy by accounting for similarities among images. Size and class balance emerge as special cases. Analyzing thousands of subsets from seven medical-imaging datasets showed that the best correlates of performance were not size or class balance but A—“big alpha”—a set of sentropies interpreted as the effective number of image–class pairs in the dataset, after accounting for similarities among images. One of these, A0, explained 67% of the variance in balanced accuracy, vs. 54% for class balance and just 39% for size. The best pair of measures was size-plus-A1 (79%), which outperformed size-plus-class-balance (74%). Subsets with the largest A0 performed up to 16% better than those with the largest size (median improvement, 8%). We propose developing methods to maximize A as a way to improve deep learning performance in medical imaging.

IPC Classification

Keywords