A method and a device for training a machine learning model to construct a digital image depicting a sample are disclosed. The model is trained using a training set of images captured with a low numerical aperture (NA) objective. This set consists of multiple images of a sample, each taken under a different, structured illumination pattern. Concurrently, a ground truth image is captured. This is a single, high-resolution image of the same sample taken with a higher NA objective and conventional illumination. The machine learning model is then trained to use the series of low-resolution, uniquely illuminated images to construct a final image that matches the high-resolution ground truth. A microscope system that utilizes this trained model to produce high-resolution images surpasses the physical resolution limits of the objective lens used for capture.