Evaluating image upsampling strategies for downstream microscopy image classification

Microscopy images are frequently downsampled due to acquisition and computational constraints, requiring reconstruction before downstream analysis. While super-resolution (SR) is typically assessed using pixel-level fidelity metrics, its impact on deep learning (DL) model behavior remains insufficiently understood. In this work, we present a study that examines how different upsampling strategies affect image quality and classification performance. Using the BloodMNIST dataset, we construct matched 224×224 datasets from 64×64 images via bicubic interpolation, SwinIR Classical, and SwinIR RealGAN DL SR models, alongside the original 224 ground-truth images. We evaluate reconstruction quality using the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR) scores and assess downstream classification performance using ResNet-50 and Vision Transformer models, with accuracy, macro-F1 score, and a confidence-aware metric, the area under the receiver operating curve for successful prediction (AUPR Success). Our results demonstrate that bicubic interpolation significantly degrades classification performance, whereas SR methods can recover class-relevant information, even better than the ground-truth data. These findings emphasize the importance of confidence-aware evaluation and unambiguous reporting of reconstruction pipelines in microscopy-based DL studies.