Pathology: GigaPath-Flash and GigaTIME-Flash 50× More Subtle

Two new open models promise to drastically reduce the cost of whole slide analyses. Distilled from GigaPath and GigaTIME, they maintain high performance but are not intended for clinical use and still require broader validation.

Quantified time and memory gains, but validation still needs to be expanded

GigaPath-Flash reduces computation by about 50 times while retaining 97% of GigaPath's predictive performance. GigaTIME-Flash is approximately 6 times faster and consumes about 8 times less memory than GigaTIME, while also improving prediction quality. For estimation on a single NVIDIA A100 GPU, generating virtual mIFs with GigaTIME-Flash takes about 2 GPU hours for 1,000 slides, 7 GPU days for 100,000, and 70 GPU days for 1,000,000, compared to about 7 hours, 30 days, and 300 days for GigaTIME. These estimates assume nearly 10,000 tiles per slide and a batch size of 128, and real-time performance depends notably on slide size, tiling resolution, and hardware. The authors describe this research publication as early, with current evaluations limited in benchmarks and cohorts, and indicate that broader validation across tasks, scanners, and populations will be necessary before any clinical application. They also note that performance may vary depending on datasets, scanners, institutions, populations, and use cases.

What Flash models enable at the cohort scale

The Flash range enhances GigaPath and GigaTIME by significantly increasing their efficiency, allowing for the processing of larger cohorts and conducting more experiments. This aims to promote large-scale research in pathology to study disease biology, biomarkers, and clinical outcomes from cancer datasets. The authors emphasize that population-scale research requires repeating phases of feature extraction, statistical analyses, and validations on subgroups, operations currently hindered by computational costs. They advocate for the need for models that can be applied repeatedly and affordably to broaden the range of questions researchers can realistically ask.

GigaPath-Flash: distillation, LongNet, and minimal inference cost

GigaPath-Flash combines a ViT-S tile encoder with 22 million parameters with a LongNet slide encoder totaling 21 million parameters. The former is derived from a distillation of GigaPath ViT-g, allowing the transfer of capabilities from a one-billion-parameter model to a much lighter version, while the latter provides global contextualization of embeddings through dilated attention, whose complexity is linear relative to the number of tiles. On the PANDA (prostate classification) and EBRAINS (brain tumor subtyping) test sets, this model shows the lowest inference cost among methods pre-trained on whole slides while remaining performant, at about 3% off the original GigaPath model. It requires approximately 50 times less computational resources to achieve 97% of GigaPath's predictive performance.

GigaTIME-Flash: from CNN to ViT-S, LoRA, and out-of-distribution gains

GigaTIME-Flash replaces the historical CNN architecture with the ViT-S encoder from GigaPath-Flash, supported by a lightweight convolutional decoder for translating H&E to mIF. Fine-tuning is performed via LoRA adapters while largely keeping the weights of the pre-trained encoder fixed. On in-distribution and out-of-distribution cohorts encompassing brain, breast, colon, and lung cancers, the prediction quality of spatial proteins equals or exceeds that of GigaTIME, with particularly marked gains out of distribution. The model is reported to be about 6 times faster and 8 times less memory-intensive, while improving the prediction of mIF protein channels.

The context: rich but massive data in histopathology

Each biopsy generates a whole slide describing cellular morphology at a subcellular resolution, and hospitals produce millions of them annually. These images contain useful information for diagnosis, prognosis, therapeutic selection, and studying the tumor microenvironment. However, they are very large, often exceeding gigapixel resolution, and inference requires processing thousands of tiles per slide. When research involves tens of thousands of patients, the computational cost rises rapidly and becomes a limiting factor.

Origin of the models, open-source code, and usage restrictions

GigaPath was pre-trained on real-world histopathology data from Providence and learns whole slide representations that account for both local cellular patterns and global architecture. GigaTIME, trained on 40 million cells, has been applied to over 14,000 patients and highlighted more than 1,200 statistically significant associations between immune states and clinical biomarkers. The Flash iterations are released with open weights under the Apache 2.0 license, with code and models available on HuggingFace. These models are strictly research artifacts and are not intended or validated for clinical use. The work is a collaboration between Microsoft Research, the University of Washington, and Providence, and the authors invite the community to evaluate the models and share feedback, anticipating scientific uses beyond their initial experiments.