Deep convolutional neural networks combine Raman spectral signature of serum for prostate cancer bone metastases screening

Prostate cancer most frequently metastasizes to bone, resulting in abnormal bone metabolism and the release of components into the blood stream. Here, we evaluated the capacity of convolutional neural networks (CNNs) to use Raman data for screening of prostate cancer bone metastases. We used label-free surface-enhanced Raman spectroscopy (SERS) to collect 1281 serum Raman spectra from 427 patients with prostate cancer, and then we constructed a CNN based on LetNet-5 to recognize prostate cancer patients with bone metastases. We then used 5-fold cross-validation method to train and test the CNN model and evaluated its actual performance. Our CNN model for bone metastases detection revealed a mean training accuracy of 99.51% ± 0.23%, mean testing accuracy of 81.70% ± 2.83%, mean testing sensitivity of 80.63% ± 5.07%, and mean testing specificity of 82.82% ± 2.94%.