Resolution enhancement and target segmentation of medical images based on the frequency-domain information in deep learning.

Cancer has become a major threat to human health, with precise cellular morphology analysis critical for diagnosis and grading. Deep learning-based automatic cell segmentation is emerging as a key tool in computer-aided pathology. However, distortion and blur in digital pathology images often degrade segmentation model performance. To address this, we propose the frequency-domain resolution network, which maps images to the frequency domain, processes amplitude and phase information independently, and employs a fusion strategy to restore clear images. This approach surpasses traditional spatial-domain methods, enhancing image detail and structural feature restoration. Using these generated images, we perform nucleus extraction and segmentation, incorporating a pyramid pooling module to optimize accuracy. Experimental results show our method achieves superior resolution-enhancement reconstruction and cell segmentation, demonstrating significant potential and academic value.