JU INSIGHT Stimulated Raman Histology and Artificial Intelligence: Near-Real-Time Prostate Biopsy Interpretation

By: M. P. Mannas, MD, University of British Columbia, Vancouver, Canada, Vancouver Prostate Centre, British Columbia, Canada, NYU Langone Health, New York, New York; F. M. Deng, MD, NYU Langone Health, New York, New York; A. Ion-Margineanu, PhD, Invenio Imaging, Santa Clara, California; D. Jones, MD, NYU Langone Health, New York, New York; D. Hoskoppal, MD, NYU Langone Health, New York, New York; J. Melamed, MD, NYU Langone Health, New York, New York; S. Pastore, MD, Invenio Imaging, Santa Clara, California; C. Freudiger, PhD Invenio Imaging, Santa Clara, California; D. A. Orringer, MD, NYU Langone Health, New York, New York; S. S. Taneja, MD, NYU Langone Health, New York, New York | Posted on: 18 Mar 2024

Mannas MP, Deng FM, Ion-Margineanu A, et al. Stimulated Raman histology interpretation by artificial intelligence provides near-real-time pathologic feedback for unprocessed prostate biopsies. J Urol. 2024;211(3):384-391.

Study Need and Importance

Prostate cancer diagnosis has a long history that relies on time-consuming and sometimes inaccurate methodologies. The emergence of stimulated Raman histology (SRH) coupled with artificial intelligence (AI) presents a novel opportunity for real-time, accurate diagnosis of prostate cancer. This study evaluated the effectiveness of an AI convolutional neural network in interpreting prostate biopsy images created by SRH, offering a potential paradigm shift in prostate cancer diagnostics and patient care.

What We Found

Our study revealed that SRH can generate high-quality AI-interpretable images of fresh, unstained prostate biopsies within 2 to 2.75 minutes. At the gland level the AI showed an accuracy of 98.6% (Table). The AI demonstrated remarkable accuracy (96.5%), sensitivity (96.3%), and specificity (96.6%) in identifying prostate cancer in whole biopsies as well (Table).

Table. Results of Prostate Cancer Identification on Stimulated Raman Histology With Convolutional Neural Network at Full Scan Speed and 4× Increased Scan Speed

	AUC	Accuracy training patches	Accuracy validation patches	Ex vivo whole biopsy accuracy	In vivo whole biopsy accuracy	Ex vivo, in vivo whole biopsy combined sensitivity	Ex vivo, in vivo whole biopsy combined specificity	Ex vivo, in vivo whole biopsy combined accuracy
Full scan speed	99	99.6%	98.6%	98.3%	94.4%	96.3%	96.6%	96.5%
4× increased scan speed	99.5	N/A	93.8%	96.6%	94.4%	94.6%	96.5%	95.6%
Abbreviations: AUC, area under the curve; N/A, not available. The algorithm’s performance was evaluated on prostate biopsies obtained from various sources, including training patches (representing 96% of the total training patches), validation patches (representing 4% of the total patches), ex vivo biopsies from radical prostatectomy specimens, and in vivo biopsies.

Limitations

While promising, this technology’s current iteration cannot yet assign tumor grades. Furthermore, the study’s generalizability may be limited due to the testing in a single center on patients with confirmed prostate cancer. Future studies should focus on validating the accuracy of SRH interpretation with AI in diverse patient populations in a multicenter study and assessing its ability to grade prostate cancer.

Interpretation for Patient Care

Implementing SRH interpreted by AI in clinical practice could significantly impact patient care. It offers the potential for real-time tissue evaluation, improved targeting during biopsy procedures, and reduced rates of false-negative diagnostic procedures. Additionally, this technology could be instrumental in identifying positive surgical margins during radical prostatectomy or focal therapy, enhancing oncologic outcomes.