Poster M3 - Leveraging non-standard radiomics for outcome prediction in brain metastases treated with Personalized Ultra-fractionated Stereotactic Adaptive Radiotherapy (PULSAR)

Introduction: Personalized Ultra-fractionated Stereotactic Adaptive Radiotherapy (PULSAR) represents a paradigm shift in radiotherapy pioneered at UTSW. This study proposes a novel non-standard radiomics approach, extracting only gradient-related features within the peritumoral region, to predict outcomes in brain metastases treated with PULSAR. This approach surpasses conventional intratumoral radiomics by encompassing the peritumoral region and acknowledging its potential impact on treatment outcomes. By integrating these gradient-based features, our goal is to offer a more streamlined and interpretable feature set that can uncover deeper insights into treatment response and tumor evolution.

Materials and

Methods: A retrospective analysis of 39 patients with 69 brain metastases was conducted. For each lesion, 90 distinct features were extracted, derived from three gradient measures (gradient magnitude, radial gradient, radial gradient deviation), calculated across ten regions (core, whole margin, and eight octant margins), each characterized by three statistical metrics (mean, standard deviation, coefficient of variation). For the initial MRI datasets, we studied five scenarios: 1) core, 2) margin, 3) core/margin ratio, 4) octant margins, and 5) core+margin+octant margins, denoted by 1A to 1E, respectively. Specifically, we derived 9 features solely from the core region (1A) and solely from the entire margin (1B). Scenario 1C used the ratio of features between 1A and 1B. In scenario 1D, we selected 9 features from the initial 72 features (eight octants, each with its own nine features); likewise, we derived 9 features for scenario 1E from all 90 features (core, margin, eight octant margins). The above analysis was repeated based on the second MRI images (2A to 2E) and the delta mode (3A to 3E). In the latter, we performed arithmetic operations (subtraction) on each feature obtained between the first MRI and the second MRI. Finally, 15 individual models and an ensemble feature selection (EFS) model were constructed using support vector machines (SVM). The models aimed to classify lesions based on a volume reduction threshold of 20% at follow-up. Performance metrics, including sensitivity, specificity, accuracy, precision, F1 score, and area under the curve (AUC), were assessed.

Results, Conclusions, and Discussions:

Results: Overall, models utilizing selected features from octant margins (strategy D) and all regions (strategy E) outperform those based on core, whole margin, or ratio features (strategies A, B, and C). All pre-treatment models(Case1) were comparable with intra-treatment models(Case 2); however, delta models(Case 3) derived from delta features from pre-treatment and intra-treatment performed relatively poorly. The Case 1E model, using pretreatment features selected from all regions, stands out as the best performance among 15 individual models. The EFS model, leveraging features from models 1E, 2E, and 3E, demonstrates superior performance, surpassing all individual models with an AUC of 0.995, accuracy of 0.974, and F1 score of 0.940. The selected features, primarily from octant margin regions, underscore the importance of these areas in predicting treatment response.

Discussion:
Our study introduces a non-standard radiomics modeling approach within the PULSAR framework to predict treatment outcomes for brain metastases (BM), focusing on gradient-based features from intratumoral and peritumoral regions. Gradient Magnitude (GM), Radial Gradient (RG), and Radial Deviation (RD) offer insights into tumor boundaries and growth characteristics. Statistical metrics like mean, STD, and CV provide a comprehensive understanding of intensity variability. We found that features from octant margins are more predictive than those from core or entire margin regions, reflecting spatial heterogeneity in tumor progression. The EFS model, combining features from pretreatment, intratreatment, and delta scenarios, showed the best performance, indicating significant feature interactions. However, delta features were less predictive due to heterogeneous changes in specific octant regions. Our approach may be less suitable for tumors with infiltrative growth patterns, like GBM, where margins are unclear and difficult to define. Further investigation is needed to understand why specific octant regions are particularly important and to comprehensively compare performance with conventional methods like Pyradiomics in the PULSAR cohort.

Conclusion: Our study presents a novel approach concentrating on the peritumoral region, extracting gradient-related features such as radial gradient and radial gradient deviation. It not only reveals the directional characteristics of tumor response but also contributes to improving model interpretability.

Acknowledgements (Optional):