AI differentiates radionecrosis from true progression in brain metastasis upon stereotactic radiosurgery: analysis of 124 histologically assessed lesions

Abstract Background Differentiating radionecrosis from neoplastic progression after stereotactic radiosurgery (SRS) for brain metastases is a diagnostic challenge. Previous studies have often been limited by datasets lacking histologically confirmed diagnoses. This study aimed to develop automated models for distinguishing radionecrosis from disease progression on brain MRI, utilizing cases with definitive histopathological confirmation. Methods This multi-center retrospective study included patients who underwent surgical resection for suspected brain metastasis progression after SRS. Presurgical FLAIR and post-contrast T1 (T1w-ce) were segmented using a convolutional neural network (CNN) and compared with manual segmentation by means of Dice score. Radiomics features were extracted from each lesion, and a Random Forest model was trained on 70% of the internal dataset and evaluated on the remaining 30% and the complete external dataset. A 3DResNet-CNN was trained on the same split dataset. Validation was performed on the external dataset. Post-surgical histology was available for all cases. Results 124 brain metastases were included (104 from center 1 and 20 from center 2). Sole radionecrosis was histologically detected in 34 cases (27.4%). In the internal dataset, univariate and multivariate analysis identified 131 significantly different radiomics features, including GLDM_DNUN and GLDM_SDE within the enhancing area on the T1w-ce. On the external test dataset, the Random Forest model and the 3DResNet-CNN yielded accurate results in terms of accuracy (80.0%, 85.0%), AUROC (0.830, 0.893) and sensitivity (92.8%, 100%) in radionecrosis prediction, respectively. Conclusion Artificial intelligence could be employed to differentiate between radionecrosis and brain metastasis progression upon SRS, potentially reducing unnecessary surgical interventions.