Skip to main content
Fig. 2 | Acta Neuropathologica Communications

Fig. 2

From: Novel rapid intraoperative qualitative tumor detection by a residual convolutional neural network using label-free stimulated Raman scattering microscopy

Fig. 2

Stepwise semantic segmentation of SRH images for regions with tumor, non-tumor, and low quality. SRH images on the left side are shown before segmentation. In the middle probability, heatmaps are demonstrated for each output P (tumor, non-tumor, low quality). Using a sliding window algorithm, smaller parts in the SRH images created a probability distribution for each output. It is a function of neighboring overlapping patch predictions to generate a smoother overall heatmap after summing each part of the SRH images. Each heatmap is RGB color-coded as an overlay on the SRH image. Demonstration of prediction heatmap (A) with tumor (red) and low quality (blue) regions out of a Non-Hodgkin lymphoma specimen (* and + with atypical cell components), B with corresponding outputs classes from a non-small cell lung cancer brain metastasis specimen, C as well as only tumor (red) and non-tumor (green) prediction from an IDH-wildtype glioblastoma specimen (the arrow demonstrates the infiltrative tumor character). Scale bars = 100 μm

Back to article page