This study illustrated that geometric features derived from CE T1-weighted MR images at the first follow-up, approximately three months post-treatment, can serve as predictive indicators for recurrence in BMs following SRT. These geometric features encompassed total volume, necrotic volume, the presence or absence of necrosis, the CE rim width, and the ratio between volumes before and after irradiation.
We found that radiation therapy makes the surface of BMs more irregular, irrespective of their size. For lesions larger than 3 cm3, surface regularity in pre-treatment MR images, was found to be a prognostic factor. However, unlike other cancers, our finding for BMs was that regular lesions had worse prognosis. It is well known that more irregular melanoma lesions have worse prognosis17. Similar results have been found for other kinds of brain tumor, such as glioblastoma (GBM)6, meningiomas18. The same happens for prostate cancer19, lung cancer nodules20,21. It is somehow surprising that the more regular are the BMs, the worse the prognosis, which is consistent with our earlier finding that SRT transforms BMs into more irregular lesions, and it is assumed that it also enhances the patient’s prognosis. It would be interesting to study if the same property holds for other organs metastases. Understanding why this happens mechanistically would require further investigation. Perhaps irregular primary tumors, living in their host tissue, are indicators of a mesenchymal phenotype, while metastatic tumors achieve a better screening from the non-host immune system by adopting more compact shapes. This is a very interesting topic that deserves further investigation.
While our study did not reveal statistically significant differences (p = 0.051), large BMs at baseline tended to be associated with a poorer prognosis. Notably, an examination of total volume at the first follow-up after radiation treatment suggests that it could serve as a potential prognostic factor, along with the ratio between volumes before and after irradiation. It's worth highlighting that previous studies reported no significant association between tumor volume change at 6 or 12 weeks post- SRT and overall survival 22. However, these studies utilized 1–3 mm slice thickness in MR images, whereas our dataset exclusively consisted on thickness less than 2 mm. Besides, the use of overall survival to evaluate BMs may be inappropriate because metastatic patients may die from a variety of causes, including systemic disease, intracranial progression, or a combination of both. It has been proved that, in the context of BM patients, overall survival is influenced not only by intracranial control23, but also strongly influenced by the status of extracranial disease24,25.
Previous studies on the predictive/prognostic value of necrotic volume in BMs were purely qualitative, taking into account either just the presence or absence of necrosis10, or semiquantitative among three categories: absence of necrosis, less than 50%, and more than 50%11. Here, the qualitative study was repeated, by dividing the BMs in our cohort into the subgroups with and without necrotic tissue. Current results agree with those of earlier works26,27. However, a quantitative analysis was also carried out, which allowed for an improved distinction between subgroups. The amount of necrotic volume after treatment was a better predictor than the presence or absence of necrosis following SRT. A necrotic volume less than 0.1 cm3 is a predictor of good response.
The CE rim width was assessed in previous studies on morphological features of GBM to assess the relationship between total and necrotic volumes6. An analysis of the CE rim width revealed that in the case of BMs, the broader the rim, the better the prognosis. This finding contrasts with what had previously been discovered about GBM6, as happens with the surface regularly, where more regular tumors are linked to a better prognosis.
Previous radiomic studies have found predictor variables such as age and CE-T1w-based kurtosis28, or an improvement in the classification when adding features such as number of metastases, primary tumor site or sphericity to the clinical variables29. Other studies used hundreds of features which are not easy to interpret30,31 and are susceptible to over fitting, among other issues32. A recent study33 showed that employing two different platforms to extract radiomic features from the same images resulted in inconsistencies and contradicting conclusions what may reflect the fact that most radiomic features are not robust34.
A recent study developed a method to classify post SRS lesions as either progression or non-progression35. They used data from two centers (n = 123 and n = 117) and used the maximum diameter in 3 perpendicular directions to evaluate total volume, with an increase of more than 25% indicating progression. The best classification achieved an AUC = 0.80. Compared to our study, we used data from five institutions and employed similar progression criteria. However, we also took into account the time to progression.
The most innovative part of our study relied on morphological measurements obtained from standard postcontrast T1-weighted MRIs. The computation of such variables could be seen as a time-consuming process. However, volumetric evaluation of BMs has been shown to substantially improve the assessment of BM response to treatments compared with one-dimensional measurements15, what may suggest the need to incorporate those metrics to the clinical routine. The continuous improvement of AI-based tools enabled by the increased availability of BM datasets36–38 will probably lead to reliable fully automatic segmentation tools in the near future, thus speeding up the process.
This study had several strengths, the first of which is the careful lesion segmentation process. The same expert conducted each segmentation semi-automatically, and all were verified by a radiologist. Another strength was the multi-center approach of the study, which included lesions from five different institutions. Only morphological features with straightforward interpretations directly obtained from segmentations were used. A recent review on machine learning imaging biomarkers in neuro-oncology 39 concluded that these techniques do not yet generally outperform those using conventional statistical techniques. The review emphasized the need for larger datasets to facilitate a more comprehensive evaluation.
Our study had a number of limitations. To describe the characteristics of the tumor, only post-contrast T1-weighted MR images were used. Future research may incorporate additional imaging sequences. Furthermore, due to the retrospective nature of the study, the data was not collected in a predesigned way, and in some cases, significant data were missing. For instance, there was a significant lack of molecular markers beyond the specific tumor histology or adjuvant/concurrent immunotherapy or other therapies. Finally, the study was conducted by assessing each BM; future studies may take into consideration patient-by-patient assessment while accounting for all of their BMs.