OpenMidnight: Closing the Performance Gap Between Open and Proprietary Pathology Foundation Models

Abstract

Pathology foundation models offer transferable representations for a range of downstream tasks in computational pathology, but reliance on proprietary data, substantial compute requirements, and limited availability of training code and preprocessing pipelines hinder independent validation and further methodological development. We present OpenMidnight 1.1, an open-source replication and refinement of Midnight trained on approximately 12,000 publicly available whole-slide images from The Cancer Genome Atlas. Using an encoder initialized with pretrained DINOv2 weights, OpenMidnight 1.1 uses approximately 320 H100 GPU-hours for pathology pretraining, a 34-fold reduction relative to the compute reported for Midnight. Across 16 classification and four segmentation datasets in THUNDER, OpenMidnight 1.1 achieves the highest mean segmentation Dice (69.7) and ranks fourth among 14 models, ahead of Midnight, with performance that matches or exceeds models trained on substantially larger datasets in several tasks. Our ablation studies additionally examine how initialization, data volume, model size, and tissue composition affect downstream performance, providing empirical evidence to inform the development of future training recipes. Our release of the training code, preprocessing pipeline, and model weights lowers barriers to reproducible research and further development of competitive pathology foundation models.