2025 Poster Presentations
P350: PREDICTION OF CONSISTENCY IN PITUITARY NEUROENDOCRINE TUMORS (PITNET) USING MRI-BASED RADIOMICS AND MACHINE LEARNING: A SYSTEMATIC REVIEW AND META-ANALYSIS
Maria I Ocampo-Navia, MD; Mariana Agudelo-Arrieta, MD; Wilfran Perez-Mendez, MS; Alex Taub-Krivoy, MD; Nayeh Arana-Isaac, MS; Oscar H Feo-Lee, MD; Pontificia Universidad Javeriana
Background: Pituitary neuroendocrine tumors (PitNETs) represent approximately 16% of primary brain tumors. The tumor's consistency, whether soft or fibrous, significantly impacts surgical planning and outcomes. Radiomics shows potential for predicting this consistency and assessing surgical outcomes, although its predictive accuracy is still under investigation.
Methods: We conducted a systematic review following PRISMA guidelines to identify studies evaluating the prediction of pituitary adenoma consistency using radiomics and machine learning algorithms. A comprehensive literature search was performed using MeSH keywords in prominent databases, including PubMed, Embase, Cochrane, and Web of Science. Data extraction was carried out independently by two reviewers, and findings were synthesized through narrative analysis and comparative assessment.
Results: The database search yielded 828 studies after duplicate removal, a total of nine were included, encompassing a cohort of 947 patients with PitNETs who underwent consistency prediction through the development of machine learning algorithms and radiomics. Among the participants in which consistency was reported, 488 (66.8%) were soft and 243 (33.2%) were firm. 50.5% of the patients were male. The ground truth for consistency was mainly defined by surgical parameters and varied, however most of the publications established a soft consistency when the tumor was easily removable by aspiration or curettage, and firm when it required piecemeal resection by a microdissector or tumor forceps. MRI machines from different brands were used, the magnetic field strength was 3T in 6 publications and 1.5T in one publication, while one article used two machines with a field strength of 3 T and 1.5 T, respectively. The selected region of interest (ROI) was three-dimensional in 4 studies, two-dimensional in 4 studies, and unspecified in one. It was segmented manually in 6 articles, automatically in 2 and not reported in 1. The average number of extracted and selected features was 691.5 (range: 59-1561) and 10.5 (range: 4-27), respectively. Area under the curve for the prediction models proposed oscillated between 0.71-0.99 in the different studies. The Radiomics Quality Score total and percentage scores were 14.2 and 39.5%, respectively. Regarding the evaluation of the risk of bias through the QUADAS-2, it was assessed as unclear for all articles regarding patient selection and mainly as high for flow and timing. For the index test domain, it was rated as low for 3 studies, unclear for 2 and high for 4. And for the reference standard it was determined as low, unclear and high for 2,2, and 5 studies, respectively.
Conclusion: This systematic review provides a thorough evaluation of the use of radiomics in predicting PitNET consistency. Findings highlight the potential of radiomics to enhance preoperative planning by distinguishing between soft and firm tumors, which can significantly impact surgical outcomes. We emphasize that the ability to accurately predict tumor consistency can influence surgical approaches and reduce complications. Given the evolving nature of radiomics technology, this review aims to advance clinical decision-making and guide future research in refining predictive models for improved patient management.
.png)
.png)
