Pretreatment MRI reveals hidden breast cancer structural phenotype that predicts recurrence

This study asks a fundamental question: can a pretreatment MRI reveal something about how a breast tumor will respond that conventional clinical and genomic data cannot see? Kantha and Loew built an outcome-blind longitudinal DCE-MRI manifold from 1,000+ trajectories in the I-SPY2 trial. They discovered that the dominant axis of response geometry—a measure of how tumor structure evolves under therapy—was completely missed by the full clinical and genomic stack (age, receptor subtypes, MammaPrint, PAM50, treatment arm, tumor burden). Only when they added a single metric—baseline structural entropy derived from the pretreatment MRI—did that axis become strongly recoverable. A constrained representation mapping confirmed that this structure is intrinsic, not an artifact of post-hoc interpretation. The structural entropy signal persists through treatment, and crucially, the volumetric tumor shrinkage signal fades over time while entropy remains separated—a crossover from burden to structural persistence that conventional endpoints overlook.

Among complete responders (pathologic complete response), the study reveals a troubling finding: structurally disordered tumors can shrink more early on but remain structurally disordered, a "volumetric deception" invisible to endpoint labels. These patients may harbor residual risk despite appearing to have a perfect response. External validation across UCSF, I-SPY1, and Duke datasets confirmed that recurrence relevance depends on representation-dependent boundaries—meaning the same feature name can fail, transport, or entangle depending on extraction geometry. The authors argue that feature-name matching is insufficient; a reproducibility framework must account for representation geometry. This work opens the door to prospective validation trials where pretreatment MRI structural entropy could guide therapy intensification or de-escalation, especially in complete responders who may still be at risk.