Pathology foundation-model embeddings reveal a morphological landscape of breast lesions beyond diagnostic categories

Read the full article See related articles

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Background

Despite rapid advances in computational pathology, including foundation-model embeddings, most approaches remain classifier-based. Breast lesion pathology is an area of unmet clinicopathological need and a compelling setting for morphology-based modeling: diagnostic categories occupy overlapping morphological spectra and reflect heterogeneous pathways of tumor development, thereby posing substantial challenges for pathological diagnosis and subsequent clinical decision-making. However, approaches to delineate morphological continua across diagnostic categories and tumor development remain understudied. Here, we leveraged morphology-based foundation-model embeddings to define a morphologic state space across breast lesion categories and dissect how these embeddings could reflect a pathobiological continuum or sequence associated with tumor development.

Methods

To model breast lesion categories in continuous morphology space, we established a leakage-controlled, patient-aware workflow using the frozen foundation model Prov-GigaPath and the BReAst Carcinoma Subtyping (BRACS) dataset, a pathologist-annotated hematoxylin-and-eosin (H&E) breast lesion dataset spanning normal tissue, benign proliferative lesions, atypical lesions, ductal carcinoma in situ (DCIS) and invasive carcinoma (IC). Using the BRACS training split, we defined a train-derived principal component analysis (PCA) morphology space. We then performed two complementary analyses: normal-rooted k-nearest-neighbor (kNN)-Dijkstra graph-distance scoring to derive a trajectory-like morphology score without imposing the full ordinal diagnostic sequence; pairwise Energy distance and squared maximum mean discrepancy with a radial-basis-function kernel (RBF-MMD²) to quantify order-free distributional separation among diagnostic categories. Validation and test data were projected into the train-defined reference space to assess reproducibility.

Results

Prov-GigaPath embeddings defined a train-derived morphology space with category-associated organization. Without imposing the full diagnostic sequence, normal-rooted kNN-Dijkstra graph distances recovered a trajectory-like morphology arrangement from normal/benign to malignant lesions. Order-free Energy distance and RBF-MMD² further supported graded separation among lesion categories. Validation and test datasets preserved the pathological organization. Furthermore, morphological embeddings dissected DCIS heterogeneity associated with concurrent invasive carcinoma, without requiring additional molecular information.

Conclusions

Leveraging morphology-based foundation-model embeddings, we defined a morphologic state space across breast lesion categories, allowing nuanced positioning of breast lesions along a pathobiological continuum associated with tumor development. Our approaches pave the way for extracting biologically and clinically relevant information from routine histopathology to guide molecular testing, individualized risk assessment, and cost-efficient clinicopathological decision-making.

Highlights

  • Overlapping morphological features challenge breast diagnostic categorization

  • Foundation-model embeddings defined a morphologic state space

  • Embeddings allow nuanced positioning along morphological continua across categories

  • Embeddings dissected DCIS heterogeneity associated with concurrent invasive carcinoma

  • Latent morphological signals in H&E slides may refine individualized risk assessment

Article activity feed