For ER prediction, lesion count showed borderline significance as a covariate ( p = 0.014), suggesting a partial independent association consistent with the known tendency for ER-positive tumors to exhibit higher metastatic burden.
Although these associations did not reach statistical significance in our dataset (likely due to limited sample size), the position encoding's contribution to model performance suggests that anatomical location may contain subtle biological signals worth further investigation in larger cohorts. 5 Discussion This study presents a novel multi-modal DL framework for non-invasive prediction of hormone receptor status in breast cancer brain metastases, achieving highly competitive results through three key methodological components: three-modal complementary fusion, multi-task collaborative learning, and anatomical position encoding.