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

Low-grade astrocytomas are common, slow-growing tumors that can progress to high-grade forms if treatment is unsuccessful, with an average survival of 4.7 to 9.8 years. Five-year survival is 80% for low-grade but under 5% for high-grade with all grades proving invasive and treatment-resistant. Post-surgery radiotherapy, requiring precise dosing to protect healthy cells, is key for efficacy and minimal side effects. This research predicts the best dose of radiation therapy for patients with astrocytoma based on the analysis of images of brain patients and discriminative models. Data related to 33 patients were collected from the Mahdieh Radiation Oncology Department. These data included 2745 MRI images of subjects suffering from low-grade astrocytoma (axial, sagittal and coronal views) with a resolution of 512 × 512 pixels, along with the clinical characteristics and treatment characteristics of the patients.

In this research, two problems of regression and classification were examined. The purpose of the regression problem is to simultaneously estimate the patient’s radiotherapy sessions and their corresponding dosage, and in the classification problem, the purpose of the classification problem is to classify patients into 4 classes according to the amount of prescribed dose using past data. By combining the VIT model and the CNN network, a powerful feature extraction model from images was designed, and then regression and classification problems were solved with the help of the MLP network and SVM and Random Forest algorithms.

The best results were related to the CNN_VIT-b16 model, which was able to predict the number of radiotherapy sessions with a mean absolute error of 0.005 and an R² score of 0.993 in the problem of predicting the prescribed dose of radiotherapy with the mean absolute error of 0.0034 and an R² score of 0.998. Moreover, in the classification problem, it achieves 0.99 accuracy and 0.99 F1 score on the test data.

This study demonstrated that a hybrid CNN-ViT model could accurately predict radiotherapy dosage and session counts for low-grade astrocytoma patients using MRI images and clinical data, achieving near-perfect regression and classification performance. The model's strong results suggest it could serve as an effective decision-support tool to personalize treatment and reduce harm to healthy tissue. Despite the promising outcomes, validation on larger, more diverse datasets is needed before clinical deployment.

This research designed a diagnostic-aided model that predicts the radiotherapy plan for treating patients with astrocytoma. The radiotherapy dose and number of sessions are tailored to the individual, varying with tumor size, type, and the patient's overall health, thus complicating treatment planning. A model designed to consider these factors could aid doctors in diagnosing and treating low-grade astrocytoma. Such a model could serve as a valuable diagnostic support tool.