OpenSlideFM: A Computationally Efficient Multi-Scale Foundation Model for Computational Pathology

BackgroundComputational pathology increasingly relies on foundation models pre-trained on large-scale histopathology datasets, but existing models require substantial computational resources that limit accessibility for resource-constrained institutions. We present OpenSlideFM, a computationally efficient multi-scale foundation model that balances performance with practical deployment requirements. MethodsWe developed a dual-scale transformer architecture that simultaneously processes high-resolution (0.5 m/pixel) and low-resolution (2.0 m/pixel) features to capture both cellular morphology and tissue architecture. The model was pre-trained using self-supervised learning on 20,000 whole-slide images from The Cancer Genome Atlas spanning 31 cancer types and 10,795 patients (Table 1). We validated OpenSlideFM across four independent tasks: pan-cancer classification, lymph node metastasis detection, pathological staging, and prostate cancer grading. O_TBL View this table: org.highwire.dtl.DTLVardef@117e6b6org.highwire.dtl.DTLVardef@2be886org.highwire.dtl.DTLVardef@aeead8org.highwire.dtl.DTLVardef@1bc1897org.highwire.dtl.DTLVardef@1f28b6d_HPS_FORMAT_FIGEXP M_TBL O_FLOATNOTable 1C_FLOATNO O_TABLECAPTIONMulti-Dataset Evaluation Design C_TABLECAPTION C_TBL ResultsOpenSlideFM achieved 81.21% accuracy on 31-class pan-cancer classification, macro-AUROC of 98.65%. Multi-scale architecture significantly outperformed single-scale variants, with 2.35% absolute improvement over high-resolution alone. External validation demonstrated robust generalization: 77.4% AUROC for lymph node metastasis detection, 0.254 quadratic kappa for multi-center pathological staging, and 0.826 quadratic kappa for prostate cancer grading. The model requires only 35 million parameters and trains on a single consumer-grade workstation with NVIDIA GeForce RTX 4090 GPU (24 GB VRAM, 16-core CPU, 384 GB RAM), enabling accessible deployment compared to 300-1850 million parameters and datacenter GPU requirements for existing foundation models. ConclusionsOpenSlideFM demonstrates that computationally efficient foundation models can achieve competitive performance across diverse histopathology tasks while maintaining practical deployment feasibility. The multi-scale architecture provides complementary information from cellular and tissue levels, and the reduced computational requirements democratize access to foundation model capabilities for resource-constrained medical institutions.