Model Overview

Body CT segmentation models are evolving. Starting from abdominal multi-organ segmentation model [1]. Now the community is developing hundreds of target anatomies. In this bundle, we provide re-trained models for (3D) segmentation of 104 whole-body segments.

This model is trained using the SegResNet [3] network. The model is trained using TotalSegmentator datasets [2].

Figure source from the TotalSegmentator [2].

MONAI Label Showcase

• We highlight the use of this bundle to use and visualize in MONAI Label + 3D Slicer integration.

Data

The training set is the 104 whole-body structures from the TotalSegmentator released datasets. Users can find more details on the datasets at https://github.com/wasserth/TotalSegmentator. All rights and licenses are reserved to the original authors.

• Target: 104 structures
• Modality: CT
• Source: TotalSegmentator
• Challenge: Large volumes of structures in CT images

Preprocessing

To use the bundle, users need to download the data and merge all annotated labels into one NIFTI file. Each file contains 0-104 values, each value represents one anatomy class. We provide sample datasets and step-by-step instructions on how to get prepared:

Instruction on how to start with the prepared sample dataset:

1. Download the sample set with this link.
2. Unzip the dataset into a workspace folder.
3. There will be three sub-folders, each with several preprocessed CT volumes:
   • imagesTr: 20 samples of training scans and validation scans.
   • labelsTr: 20 samples of pre-processed label files.
   • imagesTs: 5 samples of sample testing scans.
4. Usage: users can add --dataset_dir <totalSegmentator_mergedLabel_samples> to the bundle run command to specify the data path.

Instruction on how to merge labels with the raw dataset:

• There are 104 binary masks associated with each CT scan, each mask corresponds to anatomy. These pixel-level labels are class-exclusive, users can assign each anatomy a class number then merge to a single NIFTI file as the ground truth label file. The order of anatomies can be found here.

Training Configuration

The segmentation of 104 tissues is formulated as voxel-wise multi-label segmentation. The model is optimized with the gradient descent method minimizing Dice + cross-entropy loss between the predicted mask and ground truth segmentation.

The training was performed with the following:

• GPU: 48 GB of GPU memory
• Actual Model Input: 96 x 96 x 96
• AMP: True
• Optimizer: AdamW
• Learning Rate: 1e-4
• Loss: DiceCELoss

Evaluation Configuration

The model predicts 105 channels output at the same time using softmax and argmax. It requires higher GPU memory when calculating metrics between predicted masked and ground truth. The consumption of hardware requirements, such as GPU memory is dependent on the input CT volume size.

The recommended evaluation configuration and the metrics were acquired with the following hardware:

• GPU: equal to or larger than 48 GB of GPU memory
• Model: high resolution model pre-trained at a slice thickness of 1.5 mm.

Note: there are two pre-trained models provided. The default is the high resolution model, evaluation pipeline at slice thickness of 1.5mm, users can use the lower resolution model if out of memory (OOM) occurs, which the model is pre-trained with CT scans at a slice thickness of 3.0mm.

Users can also use the inference pipeline for predicted masks, we provide detailed GPU memory consumption in the following sections.

Memory Consumption

• Dataset Manager: CacheDataset
• Data Size: 1000 3D Volumes
• Cache Rate: 0.4
• Single GPU - System RAM Usage: 83G
• Multi GPU (8 GPUs) - System RAM Usage: 666G

Memory Consumption Warning

If you face memory issues with CacheDataset, you can either switch to a regular Dataset class or lower the caching rate cache_rate in the configurations within range [0, 1] to minimize the System RAM requirements.

Input

One channel

• CT image

Output

105 channels

• Label 0: Background (everything else)
• label 1-105: Foreground classes (104)

Resource Requirements and Latency Benchmarks

GPU Consumption Warning

The model is trained with 104 classes in single instance, for predicting 104 structures, the GPU consumption can be large.

For inference pipeline, please refer to the following section for benchmarking results. Normally, a CT scans with 300 slices will take about 27G memory, if your CT is larger, please prepare larger GPU memory or use CPU for inference.

High-Resolution and Low-Resolution Models

We retrained two versions of the totalSegmentator models, following the original paper and implementation. To meet multiple demands according to computation resources and performance, we provide a 1.5 mm model and a 3.0 mm model, both models are trained with 104 foreground output channels.

In this bundle, we configured a parameter called highres, users can set it to true when using 1.5 mm model, and set it to false to use the 3.0 mm model. The high-resolution model is named model.pt by default, the low-resolution model is named model_lowres.pt.

In MONAI Label use case, users can set the parameter in 3D Slicer plugin to control which model to infer and train.

• Pretrained Checkpoints
  • 1.5 mm model: Download link
  • 3.0 mm model: Download link

Latencies and memory performance of using the bundle with MONAI Label:

Tested Image Dimension: (512, 512, 397), the slice thickness is 1.5mm in this case. After resample to 1.5 isotropic resolution, the dimension is (287, 287, 397)

1.5 mm (highres) model (Single Model with 104 foreground classes)

Benchmarking on GPU: Memory: 28.73G

• ++ Latencies => Total: 6.0277; Pre: 1.6228; Inferer: 4.1153; Invert: 0.0000; Post: 0.0897; Write: 0.1995

Benchmarking on CPU: Memory: 26G

• ++ Latencies => Total: 38.3108; Pre: 1.6643; Inferer: 30.3018; Invert: 0.0000; Post: 6.1656; Write: 0.1786

3.0 mm (lowres) model (single model with 104 foreground classes)

GPU: Memory: 5.89G

• ++ Latencies => Total: 1.9993; Pre: 1.2363; Inferer: 0.5207; Invert: 0.0000; Post: 0.0358; Write: 0.2060

CPU: Memory: 2.3G

• ++ Latencies => Total: 6.6138; Pre: 1.3192; Inferer: 3.6746; Invert: 0.0000; Post: 1.4431; Write: 0.1760

Performance

1.5 mm Model Training

Training Accuracy

Validation Dice

Please note that this bundle is non-deterministic because of the trilinear interpolation used in the network. Therefore, reproducing the training process may not get exactly the same performance. Please refer to https://pytorch.org/docs/stable/notes/randomness.html#reproducibility for more details about reproducibility.

TensorRT speedup

This bundle supports acceleration with TensorRT. The table below displays the speedup ratios observed on an A100 80G GPU.

method	torch_fp32(ms)	torch_amp(ms)	trt_fp32(ms)	trt_fp16(ms)	speedup amp	speedup fp32	speedup fp16	amp vs fp16
model computation	88.20	37.1	39.2	36.9	2.38	2.25	2.39	1.01
end2end	3717.14	2596.77	2517.29	2501.37	1.43	1.48	1.49	1.04

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_fp32 and torch_amp are for the PyTorch models with or without amp mode.
• trt_fp32 and trt_fp16 are for the TensorRT based models converted in corresponding precision.
• speedup amp, speedup fp32 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: 8.6.1+cuda12.0
• Torch-TensorRT Version: 1.4.0
• CPU Architecture: x86-64
• OS: ubuntu 20.04
• Python version:3.8.10
• CUDA version: 12.1
• GPU models and configuration: A100 80G

MONAI Bundle Commands

In addition to the Pythonic APIs, a few command line interfaces (CLI) are provided to interact with the bundle. The CLI supports flexible use cases, such as overriding configs at runtime and predefining arguments in a file.

For more details usage instructions, visit the MONAI Bundle Configuration Page.

Execute training:

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

Please note that if the default dataset path is not modified with the actual path in the bundle config files, you can also override it by using --dataset_dir:

python -m monai.bundle run --config_file configs/train.json --dataset_dir <actual dataset path>

Override the train config to execute multi-GPU training:

torchrun --standalone --nnodes=1 --nproc_per_node=2 -m monai.bundle run --config_file "['configs/train.json','configs/multi_gpu_train.json']"

Please note that the distributed training-related options depend on the actual running environment; thus, users may need to remove --standalone, modify --nnodes, or do some other necessary changes according to the machine used. For more details, please refer to pytorch's official tutorial.

Override the train config to execute evaluation with the trained model:

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

Override the train config and evaluate config to execute multi-GPU evaluation:

torchrun --standalone --nnodes=1 --nproc_per_node=2 -m monai.bundle run --config_file "['configs/train.json','configs/evaluate.json','configs/multi_gpu_evaluate.json']"

Execute inference:

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

Execute inference with Data Samples:

python -m monai.bundle run --config_file configs/inference.json --datalist "['sampledata/imagesTr/s0037.nii.gz','sampledata/imagesTr/s0038.nii.gz']"

Export checkpoint to TensorRT based models with fp32 or fp16 precision:

python -m monai.bundle trt_export --net_id network_def --filepath models/model_trt.ts --ckpt_file models/model.pt --meta_file configs/metadata.json --config_file configs/inference.json --precision <fp32/fp16> --use_trace "True"

Execute inference with the TensorRT model:

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

References

[1] Tang, Y., Gao, R., Lee, H.H., Han, S., Chen, Y., Gao, D., Nath, V., Bermudez, C., Savona, M.R., Abramson, R.G. and Bao, S., 2021. High-resolution 3D abdominal segmentation with random patch network fusion. Medical image analysis, 69, p.101894.

[2] Wasserthal, J., Meyer, M., Breit, H.C., Cyriac, J., Yang, S. and Segeroth, M., 2022. TotalSegmentator: robust segmentation of 104 anatomical structures in CT images. arXiv preprint arXiv:2208.05868.

[3] Myronenko, A., Siddiquee, M.M.R., Yang, D., He, Y. and Xu, D., 2022. Automated head and neck tumor segmentation from 3D PET/CT. arXiv preprint arXiv:2209.10809.

License

Copyright (c) MONAI Consortium

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

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

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.