HistoMet: A Pan-Cancer Deep Learning Framework for Prognostic Prediction of Metastatic Progression and Site Tropism from Primary Tumor Histopathology

Metastatic Progression remains the leading cause of cancer-related mortality, yet predicting whether a primary tumor will metastasize and where it will disseminate directly from histopathology remains a fundamental challenge. Although whole-slide images (WSIs) provide rich morphological information, prior computational pathology approaches typically address metastatic status or site prediction as isolated tasks, and do not explicitly model the clinically sequential decision process of metastatic risk assessment followed by downstream site-specific evaluation. To address this research gap, we present a decision-aware, concept-aligned MIL framework, HistoMet, for prognostic metastatic outcome prediction from primary tumor WSIs. Our proposed framework adopts a two-module prediction pipeline in which the likelihood of metastatic progression from the primary tumor is first estimated, followed by conditional prediction of metastatic site for high-risk cases. To guide representation learning and improve clinical interpretability, our framework integrates linguistically defined and data-adaptive metastatic concepts through a pretrained pathology vision-language model. We evaluate HistoMet on a multi-institutional pan-cancer cohort of 6504 patients with metastasis follow-up and site annotations. Under clinically relevant high-sensitivity screening settings (95 percent sensitivity), HistoMet significantly reduces downstream workload while maintaining high metastatic risk recall. Conditional on metastatic cases, HistoMet achieves a macro F1 of 74.6 with a standard deviation of 1.3 and a macro one-vs-rest AUC of 92.1. These results demonstrate that explicitly modeling clinical decision structure enables robust and deployable prognostic prediction of metastatic progression and site tropism directly from primary tumor histopathology.

💡 Research Summary

Metastatic progression remains the leading cause of cancer mortality, yet clinicians lack reliable tools to predict from a primary tumor’s histology whether it will metastasize and, if so, to which organ. Existing computational pathology approaches treat metastasis status and metastatic site prediction as separate, single‑step tasks and ignore the sequential clinical workflow in which a high‑sensitivity risk screen is first applied, followed by a site‑specific assessment for the screened‑positive cases. In this context, the authors introduce HistoMet, a decision‑aware, concept‑aligned multiple‑instance learning (MIL) framework that explicitly mirrors this two‑stage clinical decision process using only whole‑slide images (WSIs) of primary tumors.

The pipeline consists of two modules. Module A performs binary classification of metastatic risk. It aggregates multi‑scale (10× and 20×) patch embeddings extracted from a pretrained pathology feature extractor (Conch) using a MIL aggregator, and outputs a slide‑level probability. The operating point is set to a fixed high sensitivity (95 % in the primary experiments) to ensure that almost all future metastatic cases are captured; specificity is then used to quantify how many low‑risk primary cases can be safely excluded from downstream analysis. Module B is invoked only for the cases flagged as high‑risk by Module A. It compresses the patch‑level features into a set of learnable visual prototype tokens, aligns these tokens with semantic embeddings of clinically defined metastatic concepts (e.g., “brain metastasis”, “lymph‑node metastasis”) derived from a frozen pathology vision‑language model (a CLIP‑style model trained on histopathology image‑text pairs), and finally predicts one of four metastatic sites (brain, lymph node, liver, soft tissue). This concept‑guided alignment provides interpretability and improves performance on rare site classes.

The authors evaluated HistoMet on a multi‑institutional pan‑cancer cohort comprising 6,504 patients (6,858 WSIs) with longitudinal follow‑up and site annotations. In the first stage, 3,004 patients eventually developed metastases while 3,854 remained metastasis‑free. The second stage involved site prediction for the metastatic cases, with a highly imbalanced distribution (brain = 266, lymph node = 2,121, liver = 192, soft tissue = 425). Using five‑fold cross‑validation, the authors report that Module A consistently outperforms state‑of‑the‑art MIL baselines (ABMIL, CLAM, TransMIL) in terms of AUC and, more importantly, achieves higher specificity at the same sensitivity, thereby reducing false‑positive referrals to Module B. Calibration curves demonstrate that HistoMet’s probability estimates are well‑calibrated, supporting the use of fixed thresholds in a screening setting.

For Module B, the integration of language‑derived concepts yields a macro‑F1 of 74.6 ± 1.3 and a macro one‑vs‑rest AUC of 92.1 ± 0.9 across the four sites, surpassing all baselines. Notably, performance gains are most pronounced for the rare brain and liver metastasis classes, suggesting that semantic guidance helps the model learn discriminative patterns despite limited training examples.

When the full two‑stage pipeline is evaluated end‑to‑end under a 95 % sensitivity operating point, HistoMet achieves a five‑class accuracy of 55.8 % and a macro‑F1 of 61.2 %, outperforming CLAM and TransMIL and remaining competitive with ABMIL. Importantly, the high‑sensitivity screen filters out 23 % of cases as low‑risk, translating into a 23 % reduction in downstream workload for pathologists. At a slightly relaxed sensitivity (90 %), all methods improve modestly, but HistoMet maintains the most stable macro‑F1, highlighting its robustness to threshold variations.

Key technical contributions include: (1) explicit modeling of the clinical decision hierarchy, which mitigates error propagation and aligns model objectives with real‑world workflow; (2) the use of a pathology vision‑language model to align visual prototypes with clinically meaningful textual concepts, thereby enhancing interpretability and performance on under‑represented site classes; (3) multi‑scale feature fusion that captures both cellular‑level morphology and tissue‑level architecture, essential for detecting subtle metastatic cues.

The study acknowledges limitations such as the need for external validation on independent cohorts, the restriction to four metastatic sites, and reliance on manually crafted textual prompts. Future work could expand the site taxonomy, incorporate automated prompt generation, and explore self‑supervised pretraining on larger, more diverse histopathology datasets.

In summary, HistoMet demonstrates that a decision‑aware, concept‑aligned MIL framework can reliably predict both the likelihood of metastatic progression and the probable metastatic organ directly from primary tumor histology. By achieving high sensitivity, reducing unnecessary downstream analyses, and providing interpretable predictions, the system holds promise for integration into clinical pipelines to support early risk stratification and personalized treatment planning.