Hub

In addition to provide an end-to-end solution to Hippocampal Subfields segmentation, HSF provides a model hub that allows anyone to distribute their own segmentation ONNX models.

Those models are listed below, and directly accessible from the CLI, as in the example below:

hsf segmentation=single_accurate

Default model: bagging_accurate

Built-in Segmentation Models

Those models are coming from our original work on Hippocampal Subfields segmentation.

They are trained on T1w and T2w images coming from multiple public and private datasets. The population ranges from 4 to 80+ years old, with healthy, epileptic, MCI, Alzheimer and post-mortem patients. MRI modalities includes MRIs from low-resolution T1s (1mm iso. MPRAGE), to high-resolution coro-T2s (0.125, 0.125, 1.2mm).

MRIs are coming from multiple acquisition sites, and multiple scanners (3T, 4T and 7T).

All the models are using the same ARUnet architecture, which will be detailed in a subsequent paper.

single_fast and single_accurate models

Those are the two most basic models. Inference is done with a single model trained on 98% of the full dataset, and tested on the remaining 2%.

The accurate model does include test-time augmentation, which allows to compute more accurate and robust segmentations through a plurality vote.

Test-time augmentation also allows the computation of an uncertainty map to analyze the quality of the resulting segmentation.

bagging_fast and bagging_accurate models

Those are the two more advanced methods. Inference is done with 5 models trained on different subsets of the dataset (random sampling with replacement).

This allows to have models with different learned properties, offering a better segmentation, but a slower inference compared to the classic single_* models.

The accurate method does include test-time augmentation, which allows to compute more accurate and robust segmentations through a plurality vote.

Test-time augmentation also allows the computation of an uncertainty map to analyze the quality of the resulting segmentation.

single_sq and bagging_sq models

Those are the two most advanced methods. Inference is done with 1 or 5 models trained on different subsets of the dataset (random sampling with replacement).

This allows to have models with different learned properties, offering a better segmentation, but a slower inference compared to the classic single_* models.

Here, the models are trained with a sparsification method which allows to reduce the computational cost of the inference while retaining an optimal sub-model following the lottery ticket hypothesis.

The sq method also includes Quantization Aware Training to improve even more the efficiency of the inference.

Those methods are appropriate for recent hardware supporting efficient computations on sparse vectors. Int8 Quantization is better used on hardware supporting fast int8 matrix computations. For example, on a CPU supporting the AVX512-VNNI vector instruction set, your best bet is to use the bagging_sq segmentation method.

Test-time augmentation also allows the computation of an uncertainty map to analyze the quality of the resulting segmentation.

Third-party Segmentation Models

To date, there is no third-party segmentation model available .

If you are a researcher and you want to contribute to the model hub, please check the contributing guide. Basically, all you need to do is to export your model in ONNX format, upload it, and make a pull request with a simple YAML file defining how to download your models.