AutoSwin: learned downsampling for adapting high-resolution chest X-rays to fixed-resolution foundation models

Medical foundation models commonly operate on low-resolution inputs (e.g., 224 × 224), but the adequacy of this resolution depends on the spatial scale of the target pathology. Using chest radiographs from the VinDr ChestX dataset and the South African Medical Bureau for Occupational Diseases (MBOD), we find preliminary evidence that input resolution is a pathology-dependent bottleneck . In controlled resolution experiments, silicosis performance is strongly correlated with input resolution, improving up to 640 × 640, whereas tuberculosis performance peaks at 224 × 224 and shows no consistent trend with resolution. This descriptive divergence suggests that aggregate benchmarks can conceal the loss of fine-grained evidence needed for under-represented pathologies such as silicosis. We further evaluate AutoSwin, a trainable Hierarchical Shifted Window Transformer-based downsampling frontend that adapts high-resolution radiographs to a fixed-resolution foundation model. AutoSwin increases macro-averaged AUC from 0.5527 to 0.7437 on MBOD and from 0.8505 to 0.8873 on VinDr. These findings motivate pathology-specific resolution assessment as an important part of foundation-model design and show that learned downsampling can recover task-relevant information when retraining a high-resolution foundation model is impractical.