Fig. 2: DeepCristae outperforms state-of-the-art methods for restoring mitochondria cristae in low-resolution 2D STED images.

a Quantitative comparison of DeepCristae with conventional (Richardson-Lucy (RL)^30,31, Wiener³², SPITFIR(e)³³) and deep learning (ESRGAN⁴⁰, CARE¹⁹, RCAN³⁸, and SRResNet³⁶) image restoration methods. Metrics were computed on the test set of \({D}_{{synt}}\). Data are expressed as mean ± standard deviation. Note that all deep learning methods were trained using the same patches extracted from the training set of \({D}_{{synt}}\). Parameters used for conventional methods are indicated in Supplementary Note 2.2.2. b The image grid displays restoration results of 3 test images from \({D}_{{synt}}\) obtained with DeepCristae and two competitive deep learning methods: RCAN³⁸ and CARE¹⁹. Pixel size: 25 nm. Scale bar: 0.5 μm. White arrowheads indicate mitochondria with low contrast restored by CARE; to be compared with DeepCristae column. c Fourier Image REsolution (FIRE) was estimated using Fourier Ring Correlation⁴⁴ for 3 test images before restoration, after CARE restoration and after DeepCristae restoration. d–f Measurement of cristae widths for 155 cristae from the test set after CARE and DeepCristae restoration. Line profiles (as depicted in (d)) were fitted to a Gaussian model and FWHM was measured (Supplementary Note 1.3). d Scale bar: 0.5 μm. e Table with the number of cristae restored by CARE and DeepCristae, in comparison to the 155 observed in HR STED images, and their average width. Data are expressed as mean ± standard deviation. f Table with statistical significance from Student’s t-tests and Fisher’s tests; ns non-significant.