SLEA: a stochastic saccadic lightweight efficient attention framework with MobileNetV2 for robust and explainable 2D medical image classification

Abstract

Deep learning in medical imaging faces a persistent tension between preserving localized diagnostic information and maintaining computational efficiency. We introduce saccadic lightweight efficient attention (SLEA), a framework purpose-built for 2D medical image classification and inspired by selective visual processing in the human visual system. Rather than dividing an image into fixed patches or processing it at once, SLEA independently samples multiple localized crops from uniformly distributed spatial coordinates, with sampled locations re-drawn at every training epoch to provide spatially diverse exposure without requiring a learned fixation policy. Each glimpse is encoded through a shared ImageNet-pretrained MobileNetV2 backbone into a compact feature token; the resulting unordered token set is integrated through multi-head self-attention followed by mean pooling, capturing content-level relationships among spatially distributed regions within a minimal parameter budget. This stochastic sampling is intended to reduce dependence on fixed image locations and recurring spatial artefacts, although this regularization effect is an empirical hypothesis rather than a formally established mechanism. Experiments on ChestX-ray14, ISIC2019, and APTOS2019 show strong performance under the evaluated protocols while maintaining a compact model size of approximately 2.45 million parameters. However, the ChestX-ray14 result is substantially higher than the commonly reported benchmark range and should be interpreted cautiously; independent replication under closely matched partitions, preprocessing, optimization, and inference protocols remains necessary. Results from the COVID-19 Radiography Database are reported separately, since publicly assembled COVID-19 datasets may contain source-, acquisition-, and processing-related differences that create artificial class separability; this result is therefore excluded from the principal performance claims and is not interpreted as evidence of clinical diagnostic capability or general superiority. A multi-granularity interpretability scheme combining glimpse-level Grad-CAM, top-glimpse, and multi-glimpse visualizations provides qualitative indications of the regions associated with the model’s predictions, but these are not interpreted as exact lesion-localization maps or definitive evidence of clinically faithful reasoning.