Model Overview

Vista3D model fintuning/evaluation/inference pipeline. VISTA3D is trained using over 20 partial datasets with more complicated pipeline. To avoid confusion, we will only provide finetuning/continual learning APIs for users to finetune on their own datasets.

Continual learning

For continual learning, user can change configs/train_continual.json. More advanced users can change configurations in configs/train.json. The hyperparameters in configs/train_continual.json will overwrite ones in configs/train.json. Most hyperparameters are straighforward and user can tell based on their names. We list hyperparameters that needs to be modified.

Data

The spleen Task from the Medical Segmentation Decathalon is selected as an example to show how to continuous learning. Users can find more details on the datasets at http://medicaldecathlon.com/.

To train with other datasets, users need to provide a json data split for training and continuous learning (configs/msd_task09_spleen_folds.json is an example for reference). The data split should meet the following format ('testing' labels are optional):

{
    "training": [
        {"image": "img0001.nii.gz", "label": "label0001.nii.gz", "fold": 0},
        {"image": "img0002.nii.gz", "label": "label0002.nii.gz", "fold": 2},
        ...
     ],
    "testing": [
        {"image": "img0003.nii.gz", "label": "label0003.nii.gz"},
        {"image": "img0004.nii.gz", "label": "label0004.nii.gz"},
        ...
   ]
}

Note the data is not the absolute path to the image and label file. The actual image file will be `os.path.join(dataset_dir, data["training"][item]["image"])`, where `dataset_dir` is defined in `configs/train_continual.json`. Also 5-fold cross-validation is not required! `fold=0` is defined in train.json, which means any data item with fold==0 will be used as validation and other fold will be used for training. So if you only have 2 data, you can manually set one data to be validation by setting "fold": 0 in its datalist and the other to be training by setting "fold" to any number other than 0.

Best practice to generate data list

User can use monai to generate the 5-fold data lists. Full exampls can be found in VISTA3D open source codebase

from monai.data.utils import partition_dataset
from monai.bundle import ConfigParser
base_url = "/path_to_your_folder/"
json_name = "./your_5_folds.json"
# create matching image and label lists.
# The code to generate the lists is based on your local data structure.
# You can use glob.glob("**.nii.gz") e.t.c.
image_list = ['images/1.nii.gz', 'images/2.nii.gz', ...]
label_list = ['labels/1.nii.gz', 'labels/2.nii.gz', ...]
items = [{"image": img, "label": lab} for img, lab in zip(image_list, label_list)]
# 80% for training 20% for testing.
train_test = partition_dataset(items, ratios=[0.8, 0.2], shuffle=True, seed=0)
print(f"training: {len(train_test[0])}, testing: {len(train_test[1])}")
# num_partitions-fold split for the training set.
train_val = partition_dataset(train_test[0], num_partitions=5, shuffle=True, seed=0)
print(f"training validation folds sizes: {[len(x) for x in train_val]}")
# add the fold index to each training data.
training = []
for f, x in enumerate(train_val):
   for item in x:
      item["fold"] = f
      training.append(item)
# save json file
parser = ConfigParser({})
parser["training"] = training
parser["testing"] = train_test[1]
print(f"writing {json_name}\n\n")
if os.path.exists(json_name):
   logger.warning(f"rewrite existing datalist file: {json_name}")
ConfigParser.export_config_file(parser.config, json_name, indent=4)

Configurations

label_mappings

The core concept of label_mapping is to convert ground-truth label index of each dataset to a unified class index. For example, "Spleen" in MSD09 groundtruth will be represented by 1, while in AbdomenCT-1K it's 3. We unified a global label index (docs/labels.json) to represent all 132 classes, and create a label mapping to map those local index to this global index. So when a user is training on their own dataset, we need to know this mapping.

The current label mapping [[1, 3]] indicates that training labels' class indices 1 is mapped to the VISTA model's class 3 (format [[src_class_0, dst_class_0], [src_class_1, dst_class_1], ...]). So during inference, "3" is used to segment spleen.

Since it's finetuning, you can map your local class to any global class. If you use [[1,4]], where "4" represents pancreas, the finetuning can still work but requires more training data and epoch because the class "4" is already assigned and trained with pancreas. If you use [[1,3]], where "3" already represents spleen, the finetuning will converge much faster.

Best practice to set label_mapping

For a class that represent the same or similar class as the global index, directly map it to the global index. For example, "mouse left lung" (e.g. index 2 in the mouse dataset) can be mapped to the 28 "left lung upper lobe"(or 29 "left lung lower lobe") with [[2,28]]. After finetuning, 28 now represents "mouse left lung" and will be used for segmentation. If you want to segment 4 substructures of aorta, you can map one of the substructuress to 6 aorta and the rest to any unused classes (class > 132), [[1,6],[2,133],[3,134],[4,135]]. For a completely novel class that none of the VISTA global classes are related, directly map to unused classes (class > 132).

NOTE: Do not map to global index value >= 255. `num_classes=255` in the config only represent the maximum mapping index, while the actual output class number only depends on your label_mapping definition. The 255 value in the inference output is also used to represent 'NaN' value.

n_train_samples and n_val_samples

In train_continual.json, only n_train_samples and n_val_samples are used for training and validation. Remember to change these two values.

patch_size

The patch size parameter is defined in configs/train_continual.json: "patch_size": [128, 128, 128]. For finetuning purposes, this value needs to be changed acccording to user's task and GPU memory. Usually a larger patch_size will give better final results.

resample_to_spacing

The resample_to_spacing parameter is defined in configs/train_continual.json and it represents the resolution the model will be trained on. The 1.5,1.5,1.5 mm default is suitable for large CT organs, but for other tasks, this value should be changed to achive the optimal performance.

Advanced user: drop_label_prob and drop_point_prob (in train.json)

VISTA3D is trained to perform both automatic (class prompts) and interactive point segmentation. drop_label_prob and drop_point_prob means percentage to remove class prompts and point prompts during training respectively. If drop_point_prob=1, the model is only finetuning for automatic segmentation, while drop_label_prob=1 means only finetuning for interactive segmentation. The VISTA3D foundation model is trained with interactive only (drop_label_prob=1) and then froze the point branch and trained with fully automatic segmentation (drop_point_prob=1). In this bundle, the training is simplified by jointly training with class prompts and point prompts and both of the drop ratio is set to 0.25.

NOTE: If user doesn't use interactive segmentation, set `drop_point_prob=1` and `drop_label_prob=0` in train.json might provide a faster and easier finetuning process.

Other explanatory items

In train.json, validate[evaluator][val_head] can be auto and point. If auto, the validation results will be automatic segmentation. If point, the validation results will be sampling one positive point per object per patch. The validation scheme of combining auto and point is deprecated due to speed issue.

In train_continual.json, valid_remap is a transform that maps the groundtruth label indexes, e.g. [0,2,3,5,6] to sequential and continuous labels [0,1,2,3,4]. This is required by monai dice calculation. It is not related to mapping label index to VISTA3D defined global class index. The validation data is not mapped to the VISTA3D global class index.

label_set is used to identify the VISTA model classes for providing training prompts. val_label_set is used to identify the original training label classes for computing foreground/background mask during validation. The default configs for both variables are derived from the label_mappings config and include [0]:

"label_set": "$[0] + list(x[1] for x in @label_mappings#default)"
"val_label_set": "$[0] + list(x[0] for x in @label_mappings#default)"

Note: Please ensure the input data header is correct. The output file will use the same header as the input data, but if the input data is missing header information, MONAI will automatically provide some default values for missing values (e.g. np.eye(4) will be used if affine information is absent). This may cause a visualization misalignment depending on the visualization tool.

Commands

Single-GPU:

python -m monai.bundle run \
    --config_file="['configs/train.json','configs/train_continual.json']" --epochs=320 --learning_rate=0.00005

Multi-GPU:

torchrun --nnodes=1 --nproc_per_node=8 -m monai.bundle run \
    --config_file="['configs/train.json','configs/train_continual.json','configs/multi_gpu_train.json']" --epochs=320 --learning_rate=0.00005

Evaluation

Evaluation can be used to calculate dice scores for the model or a finetuned model. Change the ckpt_path to the checkpoint you wish to evaluate. The dice score is calculated on the original image spacing using invertd, while the dice score during finetuning is calculated on resampled space.

NOTE: Evaluation does not support point evaluation.`"validate#evaluator#hyper_kwargs#val_head` is always set to `auto`.

Single-GPU:

python -m monai.bundle run \
    --config_file="['configs/train.json','configs/train_continual.json','configs/evaluate.json']"

Multi-GPU:

torchrun --nnodes=1 --nproc_per_node=8 -m monai.bundle run \
    --config_file="['configs/train.json','configs/train_continual.json','configs/evaluate.json','configs/mgpu_evaluate.json']"

Other explanatory items

The label_mapping in evaluation.json does not include 0 because the postprocessing step performs argmax (VistaPostTransformd), and a 0 prediction would negatively impact performance. In continuous learning, however, 0 is included for validation because no argmax is performed, and validation is done channel-wise (include_background=False). Additionally, Relabeld in postprocessing is required to map label and pred back to sequential indexes like 0, 1, 2, 3, 4 for dice calculation, as they are not in one-hot format. Evaluation does not support point, but finetuning does, as it does not perform argmax.

Inference:

For inference, VISTA3d bundle requires at least one prompt for segmentation. It supports label prompt, which is the index of the class for automatic segmentation. It also supports point click prompts for binary interactive segmentation. User can provide both prompts at the same time.

All the configurations for inference is stored in inference.json, change those parameters:

input_dict

input_dict defines the image to segment and the prompt for segmentation.

"input_dict": "$[{'image': '/data/Task09_Spleen/imagesTs/spleen_15.nii.gz', 'label_prompt':[1]}]",
"input_dict": "$[{'image': '/data/Task09_Spleen/imagesTs/spleen_15.nii.gz', 'points':[[138,245,18], [271,343,27]], 'point_labels':[1,0]}]"

• The input_dict must include the key image which contain the absolute path to the nii image file, and includes prompt keys of label_prompt, points and point_labels.
• The label_prompt is a list of length B, which can perform B foreground objects segmentation, e.g. [2,3,4,5]. If B>1, Point prompts must NOT be provided.
• The points is of shape [N, 3] like [[x1,y1,z1],[x2,y2,z2],...[xN,yN,zN]], representing N point coordinates IN THE ORIGINAL IMAGE SPACE of a single foreground object. point_labels is a list of length [N] like [1,1,0,-1,...], which matches the points. 0 means background, 1 means foreground, -1 means ignoring this point. points and point_labels must pe provided together and match length.
• B must be 1 if label_prompt and points are provided together. The inferer only supports SINGLE OBJECT point click segmentatation.
• If no prompt is provided, the model will use everything_labels to segment 117 classes:

list(set([i+1 for i in range(132)]) - set([2,16,18,20,21,23,24,25,26,27,128,129,130,131,132]))

• The points together with label_prompts for "Kidney", "Lung", "Bone" (class index [2, 20, 21]) are not allowed since those prompts will be divided into sub-categories (e.g. left kidney and right kidney). Use points for the sub-categories as defined in the inference.json.
• To specify a new class for zero-shot segmentation, set the label_prompt to a value between 133 and 254. Ensure that points and point_labels are also provided; otherwise, the inference result will be a tensor of zeros.

label_prompt and label_dict

The label_dict defined in docs/labels.json has in total 132 classes. However, there are 5 we do not support and we keep them due to legacy issue. So in total VISTA3D support 127 classes.

"16, # prostate or uterus" since we already have "prostate" class,
"18, # rectum", insufficient data or dataset excluded.
"130, # liver tumor" already have hepatic tumor.
"129, # kidney mass" insufficient data or dataset excluded.
"131, # vertebrae L6", insufficient data or dataset excluded.

These 5 are excluded in the everything_labels. Another 7 tumor and vessel classes are also removed since they will overlap with other organs and make the output messy. To segment those 7 classes, we recommend users to directly set label_prompt to those indexes and avoid using them in everything_labels. For "Kidney", "Lung", "Bone" (class index [2, 20, 21]), VISTA3D did not directly use the class index for segmentation, but instead convert them to their subclass indexes as defined by subclass dict. For example, "2-Kidney" is converted to "14-Left Kidney" + "5-Right Kidney" since "2" is defined in subclasss dict.

Note: if the finetuning mapped the local user data index to global index "2, 20, 21", remove the `subclass` dict from inference.json since those values defined in `subclass` will trigger the wrong subclass segmentation.

resample_spacing

The optimal inference resample spacing should be changed according to the task. For monkey data, a high resolution of [1,1,1] showed better automatic inference results. This spacing applies to both automatic and interactive segmentation. For zero-shot interactive segmentation for non-human CTs e.g. mouse CT or even rock/stone CT, using original resolution (set resample_spacing to [-1,-1,-1]) may give better interactive results.

use_point_window

When user click a point, there is no need to perform whole image sliding window inference. Set "use_point_window" to true in the inference.json to enable this function. A window centered at the clicked points will be used for inference. All values outside of the window will set to be "NaN" unless "prev_mask" is passed to the inferer (255 is used to represent NaN). If no point click exists, this function will not be used. Notice if "use_point_window" is true and user provided point clicks, there will be obvious cut-off box artefacts.

Inference GPU benchmarks

Benchmarks on a 16GB V100 GPU with 400G system cpu memory.

Volume size at 1.5x1.5x1.5 mm	333x333x603	512x512x512	512x512x768	1024x1024x512	1024x1024x768
RunTime	1m07s	2m09s	3m25s	9m20s	killed

Commands

The bundle only provides single-gpu inference.

Single image inference

python -m monai.bundle run --config_file configs/inference.json

Batch inference for segmenting everything

python -m monai.bundle run --config_file="['configs/inference.json', 'configs/batch_inference.json']" --input_dir="/data/Task09_Spleen/imagesTr" --output_dir="./eval_task09"

configs/batch_inference.json by default runs the segment everything workflow (classes defined by everything_labels) on all (*.nii.gz) files in input_dir. This default is overridable by changing the input folder input_dir, or the input image name suffix input_suffix, or directly setting the list of filenames input_list.

Execute inference with the TensorRT model:

python -m monai.bundle run --config_file "['configs/inference.json', 'configs/inference_trt.json']"

By default, the argument head_trt_enabled is set to false in configs/inference_trt.json. This means that the class_head module of the network will not be converted into a TensorRT model. Setting this to true may accelerate the process, but there are some limitations:

Since the label_prompt will be converted into a tensor and input into the class_head module, the batch size of this input tensor will equal the length of the original label_prompt list (if no prompt is provided, the length is 117). To make the TensorRT model work on the class_head module, you should set a suitable dynamic batch size range. The maximum dynamic batch size can be configured using the argument max_prompt_size in configs/inference_trt.json. If the length of the label_prompt list exceeds max_prompt_size, the engine will fall back to using the normal PyTorch model for inference. Setting a larger max_prompt_size can cover more input cases but may require more GPU memory (the default value is 4, which requires 16 GB of GPU memory). Therefore, please set it to a reasonable value according to your actual requirements.

TroubleShoot for Out-of-Memory

• Changing patch_size to a smaller value such as "patch_size": [96, 96, 96] would reduce the training/inference memory footprint.
• Changing train_dataset_cache_rate and val_dataset_cache_rate to a smaller value like 0.1 can solve the out-of-cpu memory issue when using huge finetuning dataset.
• Set "postprocessing#transforms#0#_disabled_": false to move the postprocessing to cpu to reduce the GPU memory footprint.

TensorRT speedup

The vista3d bundle supports acceleration with TensorRT. The table below displays the speedup ratios observed on an A100 80G GPU. Please note for 32bit precision models, they are benchmarked with tf32 weight format.

method	torch_tf32(ms)	torch_amp(ms)	trt_tf32(ms)	trt_fp16(ms)	speedup amp	speedup tf32	speedup fp16	amp vs fp16
model computation	108.53	91.9	106.84	60.02	1.18	1.02	1.81	1.53
end2end	6740	5166	5242	3386	1.30	1.29	1.99	1.53

Where:

• model computation means the speedup ratio of model's inference with a random input without preprocessing and postprocessing
• end2end means run the bundle end-to-end with the TensorRT based model.
• torch_tf32 and torch_amp are for the PyTorch models with or without amp mode.
• trt_tf32 and trt_fp16 are for the TensorRT based models converted in corresponding precision.
• speedup amp, speedup tf32 and speedup fp16 are the speedup ratios of corresponding models versus the PyTorch float32 model
• amp vs fp16 is the speedup ratio between the PyTorch amp model and the TensorRT float16 based model.

This result is benchmarked under:

• TensorRT: 10.3.0+cuda12.6
• Torch-TensorRT Version: 2.4.0
• CPU Architecture: x86-64
• OS: ubuntu 20.04
• Python version:3.10.12
• CUDA version: 12.6
• GPU models and configuration: A100 80G

References

• Antonelli, M., Reinke, A., Bakas, S. et al. The Medical Segmentation Decathlon. Nat Commun 13, 4128 (2022). https://doi.org/10.1038/s41467-022-30695-9

• VISTA3D: Versatile Imaging SegmenTation and Annotation model for 3D Computed Tomography. arxiv (2024) https://arxiv.org/abs/2406.05285

License

Code License

This project includes code licensed under the Apache License 2.0. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Model Weights License

The model weights included in this project are licensed under the NCLS v1 License.

Both licenses' full texts have been combined into a single LICENSE file. Please refer to this LICENSE file for more details about the terms and conditions of both licenses.