(UroToday.com) The 2024 Society of Urologic Oncology (SUO) annual meeting held in Dallas, between December 3 and December 6, 2024, was host to the Prostate Cancer Session I. Dr. Anobel Odisho discussed AI in Prostate Cancer in the Diagnostic Setting.
Dr. Odisho began his presentation by emphasizing that AI systems are dynamic, emergent, continually learning, and somewhat imprecise. He then discussed the applications of AI in different settings for prostate cancer diagnosis, focusing on its role in prostate MRI.
Artificial intelligence in prostate MRI
The AI tasks in prostate MRI can be divided into two major groups: image quality and image interpretation. In terms of image quality, AI can assist in improving image acquisition, performing quality evaluation of MRI images, and post-processing MRI images to enhance clarity and detail. For image interpretation, AI plays a more complex role, aiding in the detection and targeting of suspicious lesions, segmenting lesions, diagnosing, and risk stratifying using PI-RADS. Additionally, AI can assist in image registration with other modalities and provide prognostic information to help decide on active surveillance.
Moreover, AI could be used in prostate MRI for segmentation of different parts of the anatomy, such as the apex and base, which helps in identifying and marking lesions. AI can also assist in identifying the urethra to ensure it is kept out of the treatment field, as demonstrated in the image below:
Dr. Odisho briefly discussed the ResNet imaging processing algorithm, which is a well-known algorithm used in AI for lesion classification. In a study utilizing 1,540 prostate MRIs from two different sites and a publicly available dataset of MRIs, the ResNet50, a convolutional neural network, was employed. Notably, the study found that ResNet50 (AUC 0.801) outperformed general radiologists (AUC 0.770) in lesion classification. However, it still fell short of the performance level of fellowship-trained radiologists (AUC 0.881).
Another study evaluating AI in prostate cancer lesion classification utilized 1,500 publicly available prostate MRIs. In this study, 62 radiologists from 45 sites in 20 countries read 400 prostate MRIs for the validation set. Additionally, 293 groups submitted classification algorithms. The top 5 algorithms were retrained on a larger dataset totaling 9,700 MRIs. The results were then ensembled and validated. They found that the AI system (AUC 0.91) outperformed the radiologists (AUC 0.86) in lesion classification, as shown below.
It is important to highlight that there are and will continue to be multiple challenges for implementing AI in prostate MRI, including data handling challenges, patient cohort diversity, clinical outcome consistency, and implementation logistics. A summary of these challenges is shown in the figure below.
AI in pathology can assist with cancer detection, cancer sub-typing or subclassification, tumor microenvironment characterization, biomarker detection to assess treatment response, and overall survival prediction.
Dr. Odisho briefly discussed Artera AI, which utilized data from 5,600 patients from Phase 3 randomized NRG trials. The dataset included clinical data such as age, log2 baseline PSA, Gleason score, T stage, and image data comprising approximately 16,000 slides. Artera AI was developed as a risk stratification tool to classify patients into low, medium, and high-risk groups. It aids in predicting response to short-term androgen deprivation therapy (ST-ADT) in patients receiving radiation therapy for intermediate-risk prostate cancer. Moreover, it has been shown to be prognostic for biochemical recurrence (BCR), distant metastasis, and prostate cancer-specific mortality (PCSM) at 5 and 10 years.1
Notably, the validation set for Artera AI was built using 20% data from each trial. It outperformed the NCCN risk classification in predicting several key outcomes: distant metastasis at 5 and 10 years, biochemical failure at 5 and 10 years, prostate cancer-specific mortality (PCSM) at 5 and 10 years, and overall survival (OS) at 10 years.
Artera AI reclassified 42% of patients compared to NCCN risk group and increased the number of patients with low risk in 7%.2
Lastly, he discussed PRISM a multimodal foundation model, which uses a much larger cohort of data (587,000) whole slide images from 195,000 specimens.
PRISM has been assessed for prostate and the AUC is nearly perfect (0.975). Dr Odisho thinks we can all use this to build upon and continue tunning AI for prostate cancer.
Dr. Odisho concluded his presentation with the following key takeaways:
- There are Different AI approaches: general, foundation models, and specific-task models.
- AI can be used for specific tasks or prognostication.
- Decision needed: Will AI models be open source or proprietary?
- Key barriers to clinical use: trust in models, technical integration, and workflow integration.
Presented by: Anobel Y. Odisho, MD, MPH, FAMIA, Associate Professor of Urology, Epidemiology/Biostatistics, and Medicine at the Division of Clinical Informatics and Digital Transformation, Dept of Medicine UCSF Health, California, USA.
Written by: Julian Chavarriaga, MD – Staff Urologic Oncologist at Cancer Treatment and Research Center (CTIC) Luis Carlos Sarmiento Angulo Foundation via Society of Urologic Oncology (SUO) Fellow at The University of Toronto. @chavarriagaj on Twitter during the 2024 Society of Urologic Oncology (SUO) annual meeting held in Dallas, between the 3rd and 6th of December, 2024.
References:- Esteva A, Feng J, van der Wal D, Huang SC, Simko JP, DeVries S, Chen E, Schaeffer EM, Morgan TM, Sun Y, Ghorbani A, Naik N, Nathawani D, Socher R, Michalski JM, Roach M 3rd, Pisansky TM, Monson JM, Naz F, Wallace J, Ferguson MJ, Bahary JP, Zou J, Lungren M, Yeung S, Ross AE; NRG Prostate Cancer AI Consortium; Sandler HM, Tran PT, Spratt DE, Pugh S, Feng FY, Mohamad O. Prostate cancer therapy personalization via multi-modal deep learning on randomized phase III clinical trials. NPJ Digit Med. 2022 Jun 8;5(1):71. doi: 10.1038/s41746-022-00613-w. Erratum in: NPJ Digit Med. 2023 Feb 22;6(1):27. doi: 10.1038/s41746-023-00769-z. PMID: 35676445; PMCID: PMC9177850.
- Jonathan David Tward et al.,Prostate Cancer Risk Stratification in NRG Oncology Phase III Randomized Trials Using Multimodal Deep Learning With Digital Histopathology. JCO Precis Oncol 8, e2400145(2024) DOI:10.1200/PO.24.00145