Longitudinal Deep Learning for Predicting Glioblastoma Recurrence: Integrating Pre- and Post-Operative MRI Dynamics

Abstract

Glioblastoma (GBM) is an aggressive brain tumor that nearly invariably returns, even after surgery and routine treatments. The majority of existing techniques use pre-operative imaging to attempt to forecast how these tumours will grow. They so fail to see what transpires following the initiation of treatment. This study presents a novel technique that learns from a sequence of MRI scans taken over time using an advanced deep learning method. The model is provided preoperative, postoperative, and follow-up scans that indicate the excised area and any remaining tumour. By combining 3D image analysis with a sequence understanding technique, the Spatio-Temporal Recurrence Prediction Network (ST-RPNet) system is able to learn how the tumour changes and where it is most likely to return. The objective is to generate customised 3D "risk maps" that display potential recurrence locations and provide an estimate of when this might occur by analysing patient data over time. Compared to earlier techniques, this produces more accurate forecasts and enables physicians to customise treatment and monitoring for every patient with glioblastoma, perhaps directing future more accurate radiation therapy. Compared to models that just use pre-treatment data, the ST-RPNet model trained on the BraTS dataset offers more precise and patient-specific predictions of recurrence location and timing.