Diatom Classification via Deep Learning using Raw Holograms

Project Description: Digital Lensless Holographic Microscopy (DLHM) is an imaging modality which allows for high-resolution phase imaging without the use of lenses, the alignment protocols as-sociated with them, or other bulky hardware. DLHM systems are compact, needing only a illumination source, and sensor. The sample is placed in between these two compo-nents near the source, with the sample’s distance from the sensor controlling the magni-fication. Modern computational resources make the reconstruction of such holograms feasible to most, although somewhat computationally expensive. Artificial intelligence (AI) has been previously used to substitute traditional reconstruction algorithms, as once these models are trained their results are nearly instantaneous. However, improving the performance of these models to provide quantitative results is an ongoing endeavor in the DLHM community. In this work we investigated the possibility of using AI to cir-cumvent the reconstruction process altogether for the case of classifying diatom samples. Rather that use these image processing AI models to reconstruct each hologram to de-termine the content’s of it’s field of view (FOV), we propose a prerequisite step of clas-sification to determine whether the contents of the hologram are worth reconstructing. We show that it is possible to train an AI model to classify based on the diffraction pat-tern captured by DLHM system. We validated our approach by comparing the perfor-mance of three typical image processing AI models: AlexNet, VGG16, and ResNet-18. These models were trained and tested on a simulated hologram dataset that was created using the dataset collected by Gunduz et al. at Eskişehir Technical University.

Links

SPIE Link

Conference Presentation PDF

Citation: B. Bogue-Jimenez, Raúl Castañeda, Carlos Trujillo, Ana Doblas. “Diatom Classification via Deep Learning using Raw Holograms captured by a Lenless Holographic System,” Proc. SPIE 12903, AI and Optical Data Sciences V, 12903-46 (31 January 2024).