Deep learning-assisted holo-tomographic flow cytometry with sparse data

The integration of holo-tomographic flow cytometry represents an innovative approach that synergistically combines the strengths of both techniques. By exploiting the self-rotation of cells to capture multi-angle projection images, this method facilitates label-free, quantitative, and isotropic reconstruction of the refractive index (RI) distribution, offering a transformative perspective for high-throughput, three-dimensional (3D) cell analysis. Nevertheless, a significant challenge persists: achieving a balance between high throughput and sufficient sampling angles to ensure accurate RI reconstruction. To address the sparse-angle limitations imposed by high-throughput conditions, we proposed a physics-inspired neural network for RI distribution reconstruction under missing-angle scenarios. Our approach employed the filtered back-projection algorithm to reconstruct an initial RI distribution as the network's input. Subsequently, a wave propagation model was used to compute the transmitted light field corresponding to the estimated RI distribution, which is compared to the experimentally measured light field to define the loss function. Through iterative training, the network refined the RI distribution until it converged to the reference reconstruction, without requiring any external training datasets, thereby enhancing the method's versatility. We validated this approach by reconstructing the RI distributions of vacuole-containing ovarian cancer cells and colon cancer cells internalizing carbon nanoparticles, using 25%, 50%, and 75% of the total acquired phase images. The Feature Similarity Index was employed to evaluate the network's performance quantitatively. By seamlessly integrating physical models with neural networks, this method introduces a novel paradigm for holo-tomographic flow cytometry, providing a pioneering solution for high-throughput 3D cell analysis.