2019/06/12 by Mason T. Chen, Chen, Mason T., Faisal Mahmood +5 · 1 citation
Medicine · Engineering · #Optical Imaging and Spectroscopy Techniques #Infrared Thermography in Medicine #Photoacoustic and Ultrasonic Imaging
paper · pdf · doi:10.48550/arxiv.1906.05360
We present a deep learning framework for wide-field, content-aware estimation\nof absorption and scattering coefficients of tissues, called Generative\nAdversarial Network Prediction of Optical Properties (GANPOP). Spatial\nfrequency domain imaging is used to obtain ground-truth optical properties from\nin vivo human hands, freshly resected human esophagectomy samples and\nhomogeneous tissue phantoms. Images of objects with either flat-field or\nstructured illumination are paired with registered optical property maps and\nare used to train conditional generative adversarial networks that estimate\noptical properties from a single input image. We benchmark this approach by\ncomparing GANPOP to a single-snapshot optical property (SSOP) technique, using\na normalized mean absolute error (NMAE) metric. In human gastrointestinal\nspecimens, GANPOP estimates both reduced scattering and absorption coefficients\nat 660 nm from a single 0.2/mm spatial frequency illumination image with 58%\nhigher accuracy than SSOP. When applied to both in vivo and ex vivo swine\ntissues, a GANPOP model trained solely on human specimens and phantoms\nestimates optical properties with approximately 43% improvement over SSOP,\nindicating adaptability to sample variety. Moreover, we demonstrate that GANPOP\nestimates optical properties from flat-field illumination images with similar\nerror to SSOP, which requires structured-illumination. Given a training set\nthat appropriately spans the target domain, GANPOP has the potential to enable\nrapid and accurate wide-field measurements of optical properties, even from\nconventional imaging systems with flat-field illumination.\n