Abstract Deep learning (DL) chest radiograph (CXR) models are often trained on downsampled images to reduce computational overhead, despite clinical workflows operating at high resolution. Previous studies have investigated the impact of input resolution on CXR classification accuracy, yet two fundamental pillars of safe and trustworthy AI, explainability and generalizability, remain underexplored. In this retrospective study, we evaluated how training image resolution affects CXR classification performance and explanation quality in internal versus external testing. We trained Convolutional Neural Networks (CNN) for disease classification on the SIIM-ACR Pneumothorax and RSNA Pneumonia datasets at six resolutions (ranging from 64\(\times\)64 to 1024\(\times\)1024) using five-fold cross-validation and evaluated models on internal and external test sets. Internal performance was high across resolutions (AUROC >0.85), but external testing showed substantially worse generalizability at lower training resolutions, with internal-to-external drops >20% versus 4.2%-10.7% at higher resolutions (512\(\times\)512 to 1024\(\times\)1024). Higher resolutions also produced more concise explanations, with the tightest saliency-map coverage at 1024\(\times\)1024 (<4%) across models and datasets, and improved explanation quality on external data (peak precision plateauing at 768\(\times\)768 for pneumothorax). Overall, training at higher CXR resolutions improved both generalizability and explainability, providing practical guidance for radiology AI design beyond internal test performance. Similar content being viewed by others Introduction Artificial intelligence (AI) has been rapidly adopted in radiology with a focus on using deep learning (DL) algorithms for the diagnosis of diseases1,2,3. Although medical images are routinely acquired at high resolutions to capture fine pathological details, DL models commonly down-sample medical images during training4,5,6 to reduce computational costs, which may limit their ability to identify subtle diseases. This practice contrasts sharply with clinical workflows, where radiologists rely on high-resolution images to make accurate diagnoses. Previous studies7,8 have shown that DL models trained using lower resolution medical images can still achieve high performance for certain disease