Multimodal convolutional neural networks based on the Raman spectra of serum and clinical features for the early diagnosis of prostate cancer

We collected surface-enhanced Raman spectroscopy (SERS) data from the serum of 729 patients with prostate cancer or benign prostatic hyperplasia (BPH), corresponding to their pathological results, and built an artificial intelligence-assisted diagnosis model based on a convolutional neural network (CNN). We then evaluated its value in diagnosing prostate cancer and predicting the Gleason score (GS) using a simple cross-validation method. Our CNN model based on the spectral data for prostate cancer diagnosis revealed an accuracy of 85.14 ± 0.39%. After adjusting the model with patient age and prostate specific antigen (PSA), the accuracy of the multimodal CNN was up to 88.55 ± 0.66%. Our multimodal CNN for distinguishing low-GS/high-GS and GS = 3 + 3/GS = 3 + 4 revealed accuracies of 68 ± 0.58% and 77 ± 0.52%, respectively.