Fig. 3

Schematic of the pseudo-HE image generation algorithm. CARS, TPEF and SHG images of the samples were acquired and combined into multimodal images using a RGB color model. The multimodal images were preprocessed as follows. First, noise was removed by median filtering and the images were down-sampled by a factor of 4. After that the uneven illumination of the single tile-scans was corrected and the contrast was adjusted. For generation of the pseudo-HE stained images, a RGB color image and two masks were calculated. To convert a multimodal image into RGB values of pseudo-HE image, a partial-least-square regression (PLS) method with 3 components was used which was trained with one image. The mask of cell nuclei was predicted by a pre-trained LDA model and was used to color cell nuclei regions within the pseudo-HE stained image in dark violet. Moreover, a background mask was estimated and used to color the background area of the pseudo-HE image in white