Stimulated Raman histology (SRH) is as a label-free optical imaging method for rapid intraoperative analysis of fresh tissue samples. The analysis of SRH images using Convolutional Neural Networks (CNN) revealed promising results to predict the main classes of neuropathological tumors including brain metastases (BM). As up to 25% of patients present with BMs as primary manifestation of previously unknown cancer, a valid timely diagnosis is essential to initiate further oncological therapy. The aim of this study was to develop a CNN for classification of the main histomorphological classes of brain metastases using SRH.

In a monocentric prospective study 303 SRH images of intraoperative tissue samples of 114 patients undergoing brain metastasis resection were obtained using a fiber‐laser‐based stimulated Raman scattering microscope. Prediction of histomorphological classes was performed using a weakly supervised attention learning based deep-learning approach (clustering-constrained-attention multiple-instance learning (CLAM)).

The typical features of seven different histomorphological classes (adenocarcinoma, squamous cell carcinoma, neuroendocrine carcinoma, melanoma, sarcoma, urothelial carcinoma, hematopoietic tumors) were visualized by SRH. In a first approach, the developed CNN for binary classification achieved an accuracy of 75% for prediction of adenocarcinoma vs. other tumor classes.

It is possible to distinguish different types of brain metastases using CNNs trained on SRH images. Further development of the approach for multi-class prediction and classification of the primary tumor is currently under investigation.