Optimizing Low-grade Astrocytoma Radiotherapy Dose Prediction through Image-based Discriminative Models

Low-grade astrocytomas are slow-growing yet invasive brain tumors that may progress to high-grade forms if treatment fails. Post-surgical radiotherapy is essential but requires precise dose planning to maximize efficacy and minimize harm to healthy tissue. This study aims to predict optimal radiotherapy dosage and number of sessions for astrocytoma patients using MRI images and clinical data.

Data from 33 patients—including 2,745 MRI images (axial, sagittal, and coronal views, 512 × 512 pixels) and clinical/treatment information—were collected from the Mahdieh Radiation Oncology Department. Regression models were developed to estimate the number of radiotherapy sessions and dosage, while classification models assigned patients to one of four dose categories based on prior data. A hybrid feature extraction model combining a Vision Transformer (ViT) and Convolutional Neural Network (CNN) was used, followed by Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Random Forest algorithms.

The CNN_VIT-b16 model delivered the best performance, predicting session numbers with a mean absolute error of 0.005 and R₂ of 0.993, and dosage with a mean absolute error of 0.0034 and R₂ of 0.998. In the classification task, it achieved an accuracy of 0.99 and an F1 score of 0.99 on the test data.

The hybrid CNN-ViT model accurately predicted radiotherapy plans based on imaging and clinical features, supporting its role as a decision-support tool for personalized treatment. Nevertheless, further validation with larger, more diverse cohorts is necessary.

This study demonstrates that a diagnostic-aided model using MRI and clinical data can effectively personalize radiotherapy planning for astrocytoma, with promise for enhancing treatment precision and safety.