Artificial intelligence models in the surgical planning of low-grade gliomas: a systematic review

TL;DR

This paper reviews how AI models help in planning surgeries for low-grade gliomas, showing promising results but highlighting the need for more research.

Contribution

A systematic review of AI/ML models in surgical planning for low-grade gliomas, comparing their performance and outcomes.

Findings

DeepMedic CNN achieved the highest DICE coefficient for tumor segmentation.

DL methods outperformed classical methods by ~10% in hyperspectral imaging.

AI models showed survival probability differences between patient groups.

Abstract

AI techniques like convolutional neural networks (CNN), deep learning (DL), and neural networks (NN) have made it easier to automatically extract important clinical data for glioma post-treatment monitoring and surgical planning.

To systematically review and analyze the role of AI/ML models in the surgical planning of LGG.

A rigorous and comprehensive systematic literature search was conducted across PubMed, Scopus, Web of Science Advance, ArXiV, and Embase (Ovid) databases from inception to July 14, 2025. Articles related to the utility of ML models in the surgical planning of LGG were included.

Our review included eight studies in both preoperative and intraoperative settings with variation in the type of AI applied, such as tumor segmentation, intraoperative neuro navigation, hyperspectral imaging, and surgical recommendation. Upon comparative analysis of mean DICE coefficients of the proposed models for segmentation, the DeepMedic CNN was found to have the highest DICE for tumor segmentation. With hyperspectral imaging, the use of MLP classifiers yields high accuracy; however, when taking into consideration the quality of tiles, DL methods outperform the classical methods by ~10%. Survival Probability using the Balanced Survival lasso-network (BSL), balanced individual treatment effect (BITES), and DeepSurv models: Difference in restricted mean survival time (DRMST) between the Consis group and In-consis group [4.75 (1.54-7.95)] for BSL, [3.81 (0.63–6.98)] for Deep Surv, and [3.76 (0.57–6.96)] for BITES.

AI/ML models have shown promising results in diagnostic and management approaches for glioma resection. Nonetheless, this is based on a small number of studies (n=8) and remain preliminary. Validating the findings in external datasets with a larger patient population would help enhance the predictive capacity of the existing models.

Figures