Fractal-based artificial class activation maps for data augmentation in whole-slide image analysis

Histopathological analysis of whole-slide images (WSIs) is central to deep learning–based cancer diagnosis. Still, the large volumes of annotated data required for training deep convolutional networks (DCNs) are costly and time-consuming to obtain. We propose a novel data augmentation method that generates artificial class activation maps (CAMs) using tissue-like fractals and patch-level scores sampled from a known probability distribution. Unlike existing approaches that depend on computationally expensive synthetic image generation, such as GANs and Diffusion Models, our method leverages prior knowledge of real data to create lightweight yet informative CAMs. Evaluation of stomach and colorectal cancer WSIs shows that incorporating artificial CAMs substantially improves classification performance, particularly when annotated data are limited. With only 25 real maps, adding our artificial CAMs increased accuracy from 62.86% to 92.35% and the AUC from 82.34% to 98.54%. These results demonstrate that our approach reduces reliance on annotated datasets while enhancing the generalization of DCNs, offering an efficient solution for cancer classification in histopathology.