Hierarchical variants of so-called deep convolutional neural networks (DCNNs) have facilitated breakthrough results for numerous pattern recognition tasks in recent years. We assess the potential of these novel whole-image classifiers for Raman-microscopy-based cytopathology. Conceptually, DCNNs facilitate a flexible combination of spectral and spatial information for classifying cellular images as healthy or cancer-affected cells. As we demonstrate, this conceptual advantage translates into practice, where DCNNs exceed the accuracy of both conventional classifiers based on pixel spectra as well as classifiers based on morphological features extracted from Raman microscopic images. Remarkably, accuracies exceeding those of all previously proposed classifiers are obtained while using only a small fraction of the spectral information provided by the dataset. Overall, our results indicate a high potential for DCNNs in medical applications of not just Raman, but also infrared microscopy.

The image displays the construction of the first layers of a deep convolutional neural network for classifying Raman microscopic images of urotheleal cells as either cancerous or normal. The input layer is obtained by the intensity images of two selected wavenumbers, indicated as red and green, respectively. Deeper layers are constructed by convolution operations indicated in the yellow box. The final layer (not shown) will distinguish the two cell types. Further details can be found in the article by Sascha D. Krauß, Raphael Roy, Hesham K. Yosef, et al. (https://doi.org/10.1002/e201800022)