• deep learning
• medical physics
• super-resolution microscopy

Super-Resolution Microscopy (SRM) is an advanced imaging technology surpassing the diffraction limit, enabling nanoscale cellular observation. In this master Thesis an Artificial Intelligence (AI) approach has been explore to overcome some of SRM acquisition limitations.

Particularly, this study is focused on the development and validation of an approach based on deep neural network to reconstruct SRM images from diffraction-limited ones. Specifically, it explores the application of the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) for SRM image generation.

Due to limited datasets, the transfer learning approach was chosen, involving fine-tuning the pre-trained ESRGAN model with microtubule images obtained through the Stochastical Optical Resolution Microscopy (STORM) technique from the Advanced Microscopy and Nanoscopy Laboratory at the University of Pisa. A public Structure Illumination Microscopy (SIM) dataset is also used to assess network performance across different acquisition techniques and image qualities.

Various fine-tuning strategies resulted in the creation of high-quality, high-resolution images, showing the adaptability of the AI-based algorithm for Super-Resolution (SR) to microscopy images. This results prove that an AI-approach is a valuable tool to expedite and improve SRM analysis. Moreover, it showcases potential applications in studying cellular radiation damage, particularly in radiotherapy, including the innovative FLASH approach.