Deployment of a real-time prostate cancer confirmation system with Raman spectroscopy: fine-tuning versus test-time adaptation of 1D CNNs.

Abstract

Significance: Prostate cancer (PCa) confirmation during needle-based procedures is limited by the lack of intraoperative diagnostic tools. Raman spectroscopy (RS), combined with classification models, offers a promising solution for real-time tissue characterization, potentially improving sampling accuracy and therapy guidance. However, such models require tissue- and organ-specific data, making deployment in studies challenging due to limited data availability.

Aim: The aim is to develop a one-dimensional convolutional neural network (1D-CNN) for real-time PCa detection using RS on prospectively collected ex vivo data, leveraging multi-organ pre-training and evaluating two domain adaptation strategies.

Approach: A ResNet-based 1D-CNN was trained for binary cancer/normal tissue classification. We implemented a pre-training strategy using retrospective RS data from brain, breast, and prostate (202 patients), along with pre-trained bacterial models, followed by efficient fine-tuning and test-time adaptation (TTA) to adapt to unseen domains.

Results: Prospective RS data were acquired using a robotic system from 10 PCa patients (two to five biopsies each). The fine-tuned model achieved 0.76 area under the receiver operating characteristic curve, 0.79 accuracy, 0.83 sensitivity, and 0.72 specificity, outperforming support vector machines. TTA improved predictions when labels were unavailable.

Conclusions: Pre-trained 1D-CNNs combined with efficient fine-tuning or TTA enable accurate PCa detection in small-cohort settings using real-time RS.