Parameter-Efficient Adaptation of a Vision Foundation Model Maps the Allometric Scaling Laws of Gastric Tumor Ecology

Background Gastric cancer (GC) remains a major public health challenge in Chile, exacerbated by a shortage of regional pathologists and population underrepresentation in global AI datasets. While pathology foundation models offer high-sensitivity screening, their deployment is limited by computational demands, domain shifts, and binary classification constraints that overlook complex tumor biology and sampling-scale artifacts. Methods We developed a parameter-efficient workflow using Low-Rank Adaptation on the MUSK vision foundation model (MUSK-LoRA). Using a cohort of 108 GC patients from a high-incidence regional Chilean hospital, we evaluated diagnostic screening performance, incorporating a refiner module to project spatial probability maps. We leveraged the specialized latent space to define 29 automated digital morphotypes via the Leiden algorithm. These units were analyzed using landscape ecology metrics and an allometric scaling framework to quantify morphotype heterogeneity while mathematically decoupling specimen sampling size from true tissue structural scaling properties. Results The adapted MUSK-LoRA ensemble achieved superior performance (ROC-AUC = 0.95, MCC = 0.81), optimizing hold-out sensitivity to 0.996, whereas zero-shot inference showed no predictive value (MCC = -0.02). Latent space partitioning identified 29 stable morphotypes (Q = 0.74) that segregated classical histological subtypes without molecular labels. Spatial mapping revealed that morphotype richness and Shannon diversity significantly correlate with T-stage and invasion depth (p < 0.001). Crucially, extracting sample-size-corrected net allometric residuals (∆LMS) successfully eliminated sampling area biases, revealing that corrected microenvironmental disorganization correlates with anatomical tumor staging (p < 0.001) independently of late-stage systemic mortality factors (p = 0.809). Conclusions Efficient local adaptation of foundation models provides a scalable path toward technological sovereignty in resource-constrained networks. By shifting from binary detection to an allometrically corrected ecological paradigm, this workflow introduces a precise framework for patient stratification tailored to regional epidemiological realities.