In the realm of oncological treatment, stereotactic radiosurgery (SRS) is key part of treatment for patients with brain metastases. However, after SRS treatment doctors sometimes see suspicious changes in the brain. These changes can either be from the treatment itself, called radiation necrosis (RN), or from cancer coming back (tumor recurrence). Knowing the difference is very important for deciding the next steps in treatment. A recently published study entitled “Leveraging radiomics and machine learning to differentiate radiation necrosis from recurrence in patients with brain metastases” co-led by collaborator Mustafa M. Basree and MIMRTL’s Chengnan Li in the Journal of Neuro-Oncology explores the use of radiomics and machine learning to better characterize RN from tumors. Other University of Wisconsin team members who contributed to the research study included Hyemin Um, Anthony H. Bui, Manlu Liu, Azam Ahmed, Pallavi Tiwari, Alan B. McMillan, and Andrew M. Baschnagel.
Study Overview
The study focused on 55 patients who underwent SRS and subsequently developed either RN or tumor recurrence. Utilizing advanced imaging techniques, image processing and radiomic feature extraction was performed on T1-weighted post-contrast (T1c), T2, and fluid-attenuated inversion recovery (FLAIR) MRI sequences. A total of 105 distinct radiomic features were extracted and analyzed. A univariate analysis to determine the significance of individual radiomic features. Interestingly, 27 features from the T1c sequence emerged as statistically significant, whereas features from the T2 and FLAIR sequences did not demonstrate significance. In addition, only the variable of immunotherapy post-SRS showed clinical relevance. For a more refined analysis, a multivariable approach was adopted using various classifiers to focus on the most discriminative features identified through feature selection. The models underwent cross-validation, with the effectiveness of each model gauged by the area under the receiver operating characteristic (ROC) curve (AUC). The best-performing model achieved an AUC of 76.2%, with a standard deviation of ±12.7%. The specificity and sensitivity were reported at 75.5% (±13.4%) and 62.3% (±19.6%) respectively, indicating a reasonable capability in differentiating RN from tumor recurrence. These findings suggest that radiomics combined with ML can enhance the diagnostic accuracy in post-SRS assessments.
Moving Forward
The promising results of this study underline the potential of radiomics and machine learning to improve diagnostic precision. “The use of these advanced technologies allows us to better interpret complex imaging data, potentially leading to more accurate diagnoses and tailored treatment plans for patients,” explains MIMRTL Director Alan McMillan, one of the study’s senior co-authors. However, the authors advocate for further validation of these findings through larger, multi-institutional studies and emphasize the importance of prospective evaluations to fully integrate this approach into clinical practice.
For More Information
The paper can be accessed here: https://doi.org/10.1007/s11060-024-04669-4