Abstract
Lens-free digital in-line holography (LDIH) provides a large field-of-view at micrometer-scale resolution, making it a promising tool for high-throughput cellular analysis. However, the complexity of diffraction images (holograms) produced by LDIH presents challenges for human interpretation and requires time-consuming computational reconstruction, often leading to artifacts and information loss. To address these issues, we present HoloNet, a novel deep learning architecture specifically designed for direct analysis of diffraction images in cellular diagnostics. Tailored to the unique characteristics of diffraction images, HoloNet captures multi-scale features, enabling it to outperform conventional convolutional neural networks in recognizing well-defined regions within complex holograms. HoloNet classifies breast cancer cell types with high precision and quantifies molecular marker intensities using raw diffraction images of cells stained with ER/PR and HER2. Additionally, HoloNet has proven effective in transfer learning applications, accurately classifying breast cancer cell lines and discovering previously unidentified subtypes through unsupervised learning. By integrating computational imaging with deep learning, HoloNet offers a robust solution to the challenges of holographic data analysis, significantly improving the accuracy and explainability of cellular diagnostics.