Author Affiliations
1Beijing Institute of Technology, MIIT Key Laboratory of Complex-Field Intelligent Sensing, Beijing, China2Beijing Institute of Technology, School of Information and Electronics and Advanced Research Institute of Multidisciplinary Sciences, Beijing, China3Hangzhou Dianzi University, School of Communication Engineering, Hangzhou, China4California Institute of Technology, Department of Electrical Engineering, Pasadena, California, United States5University of Connecticut, Department of Biomedical Engineering, Storrs, Connecticut, United States6Yangtze Delta Region Academy of Beijing Institute of Technology (Jiaxing), Jiaxing, Chinashow less
Fig. 1. Architecture of the proposed CI-CDNet. (a) Complex-domain neural network architecture. (b) Complex-domain convolution operation. (c) Multisource noise model for CI.
Fig. 2. Experimental setup and resolution test of KKR holography under only 1 ms exposure time. (a) Experimental setup. CL, collimating lens; RL, relay lens; P, polarizer; BS, beam splitter; TL, tube lens; SLM, spatial light modulator. (b) Resolution test results of different enhancing methods using Siemens star under 1 ms exposure time. The blue and red curves are the cross sections of the images, which represent pixel-wise errors. The extremely short exposure time results in a low SNR of KKR direct reconstruction. (c) Running time (ms) of different enhancing methods.
Fig. 3. Quantitative results of KKR holography using a biological sample. (a) Enhancing results of different methods using a papillary thyroid carcinoma slide. (b) Quantitative results of CI-CDNet for reducing exposure time. The result of CI-CDNet under 1 ms exposure time is close to the results of KKR recovery under 50 ms exposure time (more results can be seen in Note 5 in the Supplemental Material).
Fig. 4. Results of FPM under only 0.15 ms exposure time. (a) Experimental setup. (b) Resolution test of different enhancing methods using the USAF resolution test chart under 0.15 ms exposure time. (c) Enhancing results of the unstained blood smear under 0.15 ms exposure time. (d), (e) Quantitative results (blood smear) of amplitude and phase, respectively. The proposed CI-CDNet obtains more than 11 dB (amplitude) and 18 dB (phase) improvement on the PSNR index compared with the conventional AP method. (f) Running time (s) of different enhancing methods. AP is the baseline algorithm. Other enhancing methods are regularizers in FPM reconstruction; thus their running time includes the iteration time of data-fidelity term (Note 2 in the Supplemental Material).
Fig. 5. Results of LCP. (a) Schematic diagram of LCP system. (b), (c) Reconstructed amplitude and phase of the unstained blood smear using CI-CDNet with only 50 captured images. (d) Close-ups of three ROIs. The pseudo-color part is the phase and the gray part shows the amplitude. The GT is the result of ePIE using 900 images. (e) Results of white blood cell segmentation. (f) Results of virtual staining.
Algorithm | Data Volume | Amplitude | Phase | PSNR | SSIM | PSNR | SSIM | ePIE | 50 images | 16.38 | 0.46 | 15.14 | 0.38 | 100 images | 24.48 | 0.73 | 21.67 | 0.50 | 500 images | 30.01 | 0.93 | 26.84 | 0.79 | CI-CDNet | 50 images | 30.59 | 0.89 | 27.08 | 0.75 |
|
Table 1. Quantitative results of LCP using different data volumes. The result of CI-CDNet using 50 captured images is close to the results of ePIE using 500 images (more results can be seen in Note 5 in the Supplemental Material).