# Artificial intelligence-based volumetric measurements for longitudinal clinical assessment of treatment response in high-grade gliomas: Validation across institutional and public datasets

## TL;DR

This study evaluates an FDA-approved AI tool for measuring brain tumor volumes, finding it efficient but with limitations in accuracy compared to expert assessments.

## Contribution

Validates an FDA-cleared AI tool for HGG tumor volume measurement across public and institutional datasets.

## Key findings

- AI significantly reduced segmentation time compared to manual methods.
- Moderate agreement was found between AI-informed assessments and expert tumor board diagnoses.
- AI showed inconsistencies in T2-FLAIR-derived volumes due to over-segmentation.

## Abstract

High-grade gliomas (HGGs) require ongoing imaging to guide treatment, traditionally relying on labor-intensive and variable manual MRI measurements. While FDA-cleared artificial intelligence (AI) tools offer automated tumor volume segmentation, their clinical utility in decision-making remains understudied. This study assesses the utility and limitations of an FDA-cleared AI-based tool across public and institutional datasets, comparing its output with multidisciplinary tumor board (MDTB) assessments.

We applied the FDA-cleared, AI-based tool Neosoma HGG to quantify tumor volumes in 214 subjects from public datasets and 49 from an institutional cohort. AI-derived volumes were compared to expert manual and other AI-based measurements. Therapeutic response assessments using RANO criteria were evaluated against MDTB diagnoses. Segmentation times were analyzed using mixed-model regression.

We analyzed 1648 MRI sequences of 95 HGG patients across three datasets. Contrast-enhancing (CE) tumor volumes were consistent across AI platforms, and Neosoma HGG significantly reduced segmentation time (pre-operative: 210.5s, post-operative: 179s vs. 15 s, P < .0001). AI-informed disease state assessments showed an overall moderate agreement with MDTB diagnoses for progressive disease (k = 0.45, P < .00001). Key discrepancies arose from limitation of Neosoma HGG in distinguishing pseudo-progression from tumor progression. T2-FLAIR-derived volumes varied significantly between AI platforms (P < .001), with discordances largely due to over-segmentation beyond the tumor region.

AI-based volumetric segmentation has the potential to improve efficiency and standardization in monitoring HGG, especially for CE tumor burden. However, moderate concordance with MDTB assessments and difficulties with FLAIR imaging underscore its current limitations. AI should serve as a clinical decision support tool, with further refinement needed to improve specificity and integrate multimodal imaging data.