Fig. 1: General principle and validation of IMC-Denoise on the human bone marrow IMC dataset.

a Schematic of IMC-Denoise: (i) the DIMR algorithm: after the Anscombe transformation, the difference maps calculated from the raw image are operated to form a histogram. The outliers are detected based on this histogram and removed by a 3 × 3 median filter, iteratively. (ii) The training phase of the self-supervised DeepSNiF algorithm: In the hot pixel corrected images, several pixels are randomly selected and masked. The hot pixel corrected images before and after the masking are set as the outputs and inputs of a deep neural network, respectively. Statistics-derived I-divergence on the masked pixels combined with the Hessian norm regularization on all the pixels is set as the loss function to guarantee the optimal denoising performance. (iii) The prediction phase of the DeepSNiF algorithm: the hot pixel corrected IMC images are fed into the trained network to account for the shot noise. b The fractions of detected hot pixels by DIMR in selected channels. c DIMR removes hot pixels in DNA intercalator channel effectively. Left: Comparison of the raw and DIMR-processed images; and the difference between the images, in which Residual corresponds to the detected hot pixels. Upper right: the corresponding histograms of the raw and DIMR-processed images. Lower right: comparisons between the raw, NTHM, MTHM and DIMR processed images. d Visual inspection of DeepSNiF and other statistics-based denoising algorithms on a Collagen III-labeled IMC image. e DeepSNiF performs significantly better than other algorithms (n = 12 independent images) on denoising Collagen III-labeled IMC images in terms of STDB and CNR. f Visual inspection of DeepSNiF denoised IMC images labeled with other markers. g DeepSNiF improves the Pearson correlations between Collagen III-labeled IMC images with low and high SNR significantly (n = 10 independent images). In e and g, box center indicates median, box edges 25th and 75th percentile, and whiskers minimum and maximum percentile; P values were calculated through two-sided Wilcoxon matched-paired test (*P < 0.05, **P < 0.01, ***P < 0.001 and ns: no significance). Scale bar: a Upper: 100 μm, lower: 125 μm. c Whole region: 75 μm, sub-region 1–3: 8 μm. d 50 μm. f 45 μm. g 100 μm.