Clinical Prediction

MIST improves concordance index across all four survival cohorts, averaging +5.2% across aggregators. In CRC survival prediction, Haiku mIF embeddings improved mean C-index over VirTues but the improvement was not statistically significant. In metastatic melanoma treatment response prediction, Haiku showed large gains over VirTues. Patch-level linear probing found fused Haiku H&E and mIF embeddings outperformed unimodal and baseline representations. In CRC response prediction, Haiku was only modestly better and the reported differences were not significant. Haiku produced stronger median-risk Kaplan-Meier stratification than VirTues in CRC survival prediction. MaxMIL achieves a cross-cohort survival gain of approximately +7.2% with MIST, roughly twice the gain of attention-based aggregators such as ABMIL and CLAM. Weaker MIL aggregators benefit most from MIST on survival tasks, consistent with the hypothesis that survival is driven primarily by molecular rather than morphological state. Survival outcomes are interpreted as being driven by molecular microenvironmental composition, which H&E alone captures only indirectly.