Comparative analysis of the spatial distribution of brain metastases across several primary cancers using machine learning and deep learning models

Objective Brain metastases (BM) are associated with poor prognosis and increased mortality rates, making them a significant clinical challenge. Therefore, studying BMs can aid in developing better diagnostic tools for their early detection and monitoring. Systematic comparisons of anatomical distributions of BM from different primary cancers, however, remain largely unavailable.

Methods To test the hypothesis that anatomical BM distributions differ based on primary cancer type, we analyze the spatial coordinates of BMs for five different primary cancer types along principal component (PC) axes which optimizes their largest spread along each of the three PC axes. Data used in this analysis is taken from the International Radiosurgery Research Foundation (IRRF) and all patients underwent gamma-knife radiosurgery (GKRS) for the treatment of BMs which are labeled based on the primary cancer types Breast, Lung, Melanoma, Renal, and Colon. The dataset consists of six features including sex, age, target volume, and stereotactic Cartesian coordinates X, Y, and Z of a total of 3949 intracranial metastases. We employ PC coordinates to reduce the dimensionality of our dataset and highlight the distinctions in the anatomical spread of BMs between various cancer types. We utilized different Machine Learning (ML) algorithms: Random Forest (RF), Support Vector Machine (SVM), and TabNet Deep Learning (DL) model to establish the relationship between primary cancer diagnosis, spatial coordinates of BMs, age, and target volume.

Results Our findings demonstrate that the first principal component (PC1) exhibits a greater alignment with the Y axis, followed by the Z axis, with a minimal correlation observed with the X axis. Based on our analysis of the PC1 versus PC2 plots, we have determined that the pairs of Breast and Lung cancer, as well as Breast and Renal cancer, exhibit the most notable distinctions in their anatomical spreading patterns. In contrast, we find that the pairs of Renal and Lung cancer, as well as Lung and Melanoma, were most similar in their patterns. Our ML and DL results indicate high accuracy in distinguishing the distribution of BM for different primary cancers, with the SVM algorithm achieving a 97% accuracy rate when using a polynomial kernel and TabNet a 96% accuracy. The RF algorithm ranks PC1 as the most important discriminating feature.

Conclusions Taken together, the results demonstrate an accurate multiclass machine learning classification with respect to the distribution of brain metastases.