Pathology Foundation Models

MIST's placement at the projection layer rather than the encoder decouples molecular supervision from any specific foundation model checkpoint, allowing the prototype bank to be reused across new FM releases without retraining. Empirical performance of MIST with foundation model checkpoints other than UNI2-h is not directly reported, representing a limitation of the evaluation. Frozen pathology foundation models produce features organized along morphological axes because they are pretrained almost exclusively on H&E images and pathology reports. The evaluation uses UNI2-h, a ViT-H/14 architecture pretrained with DINOv2 producing 1,536-dimensional feature vectors, as the frozen encoder.