Abstract
P1576
Introduction:
The main goal of this systematic review is 1. To assess the prediction machine learning based radiomics features of prostate-specific membrane antigen (PSMA)-PET imaging in detection of prostate cancer (PCa), 2. Determine the most accurate model applied among machine learning (ML) models, 3. Determine the most dominant radiomics feature among all employed features.
Methods:
We conducted a literature search in online databases, including Scopus, Medline (PubMed), Web of Science, Embase (Elsevier), Cochrane library, and Google Scholar. The methodological quality of included observational studies was determined using the modified version of the Newcastle-Ottawa Scale (NOS) by two independent authors, and disagreements were resolved by discussion or a third reviewer. The results of these studies were synthesized to identify trends and patterns in the existing literature.
Results:
After a comprehensive and systematic database literature search, 479 articles were found, and 99 full-text articles were evaluated for eligibility after screening titles and abstracts and removing duplicate documents. Finally, 19 studies have determined inclusion criteria. In all the incorporated studies, 16 studies (84.21%) utilized PSMA-PET/CT as their PET modality, with the remaining 3 (15.79%) utilizing PSMA-PET/MRI as their modality. The study design of 16 studies was retrospective (84.21%) and the 3 studies were prospective (15.79%). In total, 1 study (170 patients) applied a deep learning (DL)-radiomics algorithm as well as ML-radiomics models. All 19 studies (1703 patients) applied ML-radiomics models to detect prostate cancer (PCa) with an overall accuracy rate of 81.9 (95% CI: [78.1, 85.6]), and an overall AUC rate of 0.8 (95% CI: [0.77, 0.84]). The overall sensitivity and specificity rates were 75.4 (95% CI: [69.79, 81]) and 74.3 (95% CI: [67.7, 80.9]), respectively. The most used PSMA agent was radiolabeled with Ga-68 (57.9%), followed by 18Fluorine compounds (26.31%). Several CT-based and PSMA-PET-based radiomics and numerical clinical features were investigated in studies that shape-based features such as max diameter and volume, and textural features such as entropy, contrast, and homogeneity are the dominant features among them, which correlates with patient cancer stage. Among ML-Radiomics model, the random forest classifier model had the most frequency with a mean sensitivity, specificity, and AUC of 72.5%, 77.4%, 0.76, respectively; likewise, Among DL-Radiomics model, graph attention network (GAT) model had the most frequency with a validation sensitivity, specificity, and AUC of 59%, 64%, and 0.68, respectively, and a test sensitivity, specificity, and AUC of 68%, 73%, and 0.765, respectively. The most utilized radiopharmaceutical in PET imaging was PSMA with 73.6% frequency.
Conclusions:
In conclusion, this systematic review revealed ML based on pre-therapeutic PSMA-PET/CT radiomics features, especially shape-based and textural features have a high potential to predict the detection of prostate cancer (PCa). However, among ML models, the random forest (RF) classifier model had the highest frequency, the graph attention network (GAT) model had the highest frequency in DL models, and both of them had acceptable accuracy, with the superiority of the ML model. Employing this method in practice may be beneficial in early detection and tracing more precise treatment plans in patients.
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.