Skip to content

Brainchop Model Zoo

Mohamed E. Masoud edited this page Jun 1, 2024 · 5 revisions

Users are welcome to contribute with their 3D MRI segmentation models.

Brainchop.org is designed to enhance extensibility, and we encourage end-users to train their own MeshNet models on any 3D MRI task (e.g. Tumor segmentation ) and add it to Brainchop Model zoo folder. Also, new model files (i.e. model.json and group1-shard1ofx.bin) can be directly uploaded to Brainchop UI using the "Browse" option from the "Models" list.

Interface

MeshNet deep learning architecture




Brainchop Models for v2.4.0 and above

Flash filet : A really small model, lightning fast and easy on resources. It also works great on data that is of similar quality to the HCP structural MRIs. It is called “Full brain GWM (light)” in the menu.

Thin slice : These models are of high quality but can be fragile if input T1s are highly variable and especially are far from healthy adult data.

Rough chop : This is a completely new model that we have not had on brainchop.org yet! The model may not be as refined as the 'thin slice' class on good quality data, but it produces reasonable results on a wide variety of data qualities and types, including clinical scans. It even shows some success on infant data!

Brainchop v2.2.0 Models

Subvolume GWM (failsafe) : Gray and white matter segmentation model. This model partitions T1 image into cubes of smaller 64x64x64 size and processes one at a time. This helps to overcome browser limitations but leads to longer computation and lower accuracy.

Full Brain GWM (light): Gray and white matter segmentation model. Operates on full T1 image in a single pass, but uses only 5 filters per layer. Can work on integrated graphics cards but is barely large enough to provide good accuracy. Still more accurate and faster than the Subvolume GWM (failsafe) model, but it needs at least texture size of 9159.

Full Brain GWM (large): Gray and white matter segmentation model. Operates on full T1 image in a single pass but needs a dedicated graphics card to operate since it uses 11 filters per layer. Provides the best accuracy among the provided models above and needs at least texture size of 13585.

Extract the Brain (FAST) : Extract the brain fast model operates on full T1 image in a single pass, but uses only 5 filters per layer. Can work on built-in and integrated graphics cards but is barely large enough to provide good accuracy. Still more accurate than the Extract the Brain (failsafe) model, and needs at least texture size of 9159.

Extract the Brain (High Acc) : Extract the brain high accuracy model operates on full T1 image in a single pass, but uses only 11 filters per layer. Can work on dedicated graphics cards. Still more accurate than the Extract the Brain (FAST) model.

Extract the Brain (failsafe) : This model partitions T1 image into cubes of smaller 64x64x64 size and processes one at a time. This helps to overcome browser limitations but leads to longer computation and lower accuracy than Extract the Brain (FAST) .

Compute Brain Mask (FAST) : This fast masking model operates on full T1 image in a single pass, but uses only 5 filters per layer. Can work on integrated graphics cards but is barely large enough to provide good accuracy. Still more accurate than Compute Brain Mask (failsafe) version.

Compute Brain Mask (High Acc): This masking model operates on full T1 image in a single pass, but uses 11 filters per layer. Can work on dedicated graphics cards. Still more accurate than Compute Brain Mask (FAST) version.

Compute Brain Mask (failsafe): This masking version partitions T1 image into cubes of smaller 64x64x64 size and processes one at a time. This helps to overcome browser limitations but leads to longer computation and lower accuracy than Compute Brain Mask (FAST) model.

Cortical Atlas 50: Parcellate brain cortical areas into 50 regions.It is highly recommended to use dedicated graphics card with this model.

FS aparc+aseg Atlas 104: FreeSurfer aparc+aseg atlas 104 parcellate brain areas into 104 regions. It contains a combination of the Desikan-Killiany atlas for cortical area and also segmentation of subcortical regions. It needs at least texture size of 18121.

FS aparc+aseg Atlas 104 (failsafe): FreeSurfer aparc+aseg atlas 104 parcellates brain into 104 regions. It combines Desikan-Killiany cortical atlas and subcortical regions. The model partitions T1 volume into smaller cubes for inference to overcome browser limitations but leads to longer computation and lower accuracy. It needs at least texture size of 16384.

Clone this wiki locally