Stimulated Raman Histology Interpretation by Artificial Intelligence Provides Near-Real-Time Pathologic Feedback for Unprocessed Prostate Biopsies

Abstract

Purpose: Stimulated Raman histology is an innovative technology that generates real-time, high-resolution microscopic images of unprocessed tissue, significantly reducing prostate biopsy interpretation time. This study aims to evaluate the ability for an artificial intelligence convolutional neural network to interpretate prostate biopsy histologic images created with stimulated Raman histology.

Materials and methods: Unprocessed, unlabeled prostate biopsies were prospectively imaged using a stimulated Raman histology microscope. Following stimulated Raman histology creation, the cores underwent standard pathological processing and interpretation by at least 2 genitourinary pathologists to establish a ground truth assessment. A network, trained on 303 prostate biopsies from 100 participants, was used to measure the accuracy, sensitivity, and specificity of detecting prostate cancer on stimulated Raman histology relative to conventional pathology. The performance of the artificial intelligence was evaluated on an independent 113-biopsy test set.

Results: Prostate biopsy images obtained through stimulated Raman histology can be generated within a time frame of 2 to 2.75 minutes. The artificial intelligence system achieved a rapid classification of prostate biopsies with cancer, with a potential identification time of approximately 1 minute. The artificial intelligence demonstrated an impressive accuracy of 96.5% in detecting prostate cancer. Moreover, the artificial intelligence exhibited a sensitivity of 96.3% and a specificity of 96.6%.

Conclusions: Stimulated Raman histology generates microscopic images capable of accurately identifying prostate cancer in real time, without the need for sectioning or tissue processing. These images can be interpreted by artificial intelligence, providing physicians with near-real-time pathological feedback during the diagnosis or treatment of prostate cancer.

Keywords: Raman spectroscopy; artificial intelligence; prostate cancer.

Figure 1:

Workflow of sample collection, H&E staining, conventional pathologic assessment, and AI interpretation of SRH testing.

Figure 2:

Virtual images of prostate biopsies using SRH microscope and stimulated Raman spectroscopy. The SRH microscope utilizes stimulated Raman spectroscopy to create virtual images of prostate biopsies, as demonstrated here for ISUP grade group 3 prostate cancer. The visualization of CH² bonds is achieved through the 2845cm⁻¹ Raman spectra (A), while the visualization of CH³ bonds is achieved through the 2930cm⁻¹ Raman spectra (B). The SRH image (C) is created by overlaying pseudocolors on the CH² (A) and CH³ (B) images.

Figure 3:

Workflow for stimulated Raman histology convolutional neural network (CNN) development and testing. The CNN model was developed using 1.75 million patches from 303 biopsies of 100 participants. Hematoxylin and Eosin-stained histologic diagnoses were agreed upon by two genitourinary pathologists. The CNN testing phase involved 113 SRH prostate biopsies obtained from in-vivo participants (n=54) and ex-vivo radical prostatectomy specimens (n=59).