bioRxiv 2024|Deep Learning and 3D Imaging Reveal Whole-Body Alterations in Obesity

• Title: Deep Learning and 3D Imaging Reveal Whole-Body Alterations in Obesity

• Address: https: //www.nature.com/articles/s41587-024-02528-1

• Project homepage: https: //discotechnologies.org/MouseMapper/

• GitHub repository: https: //github.com/erturklab/mouseMapper

📝Introduction|Deep Learning and 3D Imaging Reveal Whole-Body Alterations in Obesity

Background.

Research Gap and Motivation:

Objective & Contribution):

1. In order to address the above problems, the authors developed a tool called MouseMapper to address the problem of analyzing the global impact of systemic diseases such as obesity. (What was done).

2. Specifically MouseMapper is a collection of deep learning models including three main modules: neural module, immune module and tissue module. The neural module is used to segment nerves throughout the mouse body, the immune module is used to segment CD68+ immune cells, and the tissue module is used to map segmented structures to organs and tissues throughout the mouse body. Using MouseMapper, we identified structural changes in the nerve and immune cell networks over a wide range of properties at high spatial resolution. (How well did the method work and what problems did it solve)

3. Whole-body nerve segmentation in obese mice reveals structural changes in the infraorbital nerve in obesity. As well as analyzed systemic inflammatory changes (what were the results and were the findings important)

🗺️ methods

Data labeling

• Peripheral nerves

• Trigeminal nerve images

• CD68+ immune cells

• Organ tissues: fat, muscle, bone, bone marrow, etc.

Deep learning based peripheral nerve segmentation (neuromodule)|Neural Segmentation Models

Dataset Preparation

• 35 labeled patches from whole body nerves → 28 training / 7 testing

• 7 labeled patches from trigeminal nerve → 5 training / 2 testing

• i.e. test using 7 whole body nerve patches + 2 trigeminal images

• The trigeminal training images were cropped to the same size as the other patches (300³), yielding a total of 565 training patches of 300×300×300

Preprocessing

• Normalize each patch by sample percentile:
• Set 0.5% - 99.5% as the intensity range
• values outside the range are clipped

• Then perform Min-Max normalization

• Purpose: Enhance contrast, remove outliers, strengthen visibility of neural regions, and improve model results.

Model Training and Selection

• Several architectures were tried: Attention UNet, NNFormer, SwinUNETR, UNETR, VNet, 3D UNet; 128x128x128 voxels, initial learning rate of 1e-3, SGD Optimizer, learning rate decay and Binary Cross Entropy + DICE loss for 1000 epochs. The 3D UNet with the best performance is finally selected.

• Final training of 3D UNet using nnUNet pipeline using parameters: patch size = 128³, learning rate = 1e-3, optimizer = SGD, loss function = BCE + DICE, training = 2000 epoches, clDICE loss (weight 0.5) to reinforce neural topology and connectivity modeling

Deep learning based segmentation of CD68-EGFP+ cells (immune module) | Cell segmentation modeling

• Multiple network structures were tested during training: 3D UNet, VNet, Attention UNet, NNFormer, UNETR

• Trained using nnUNet pipeline

• Patch size is 128×128×128 voxels

• Uses channel-level Min-Max normalization

• Initial learning rate of 0.001, using SGD optimizer and learning rate decay

• Trained for a total of 1000 epochs

• Trained using 5-fold cross-validation

• Evaluation metrics include voxel DICE, instance DICE, and Betti Matching.

• Based on the performance of two of these metrics, 3D UNet was selected for subsequent analysis

Whole Body Organ and Tissue Segmentation (Tissue Module)|Organ and Tissue Segmentation Models

General Process

• The tissue module is divided into two tasks: organ segmentation and soft tissue (muscle, fat) segmentation.

• Organ segmentation models were based on 6 CD68 mice trained and evaluated on 4 UCHL1 mice, culminating in the use of 3D UNet models

• The soft tissue model was trained on a progressively expanded set of 387 samples, and 3D UNet was also selected as the optimal model.

• Inference process adopts a two-step approach of "organs first, then tissues", and finally realizes the segmentation of organs and tissues of the whole mouse.

Organ segmentation model training

• Five models (3D UNet, VNet, Attention UNet, NNFormer, Swin UNETR) were trained for organ segmentation using six labeled CD68-EGFP mice.

• Training was performed using the nnUNet pipeline with z-score normalization, foreground oversampling, SGD optimizer, batch size of 2, patch size of 64 × 256 × 128, and an initial learning rate of 0.01 for 1000 rounds.

• Evaluation was performed on 4 UCHL1 mice, using 5-fold cross-validation and integrating the results of 5 models. The final best performing model was 3D UNet.

Soft tissue (muscle and fat) segmentation model training

• Expanded training set to 387 samples using inference + manual correction

• Trained 3D UNet, VNet, Attention UNet and UNETR using 5-fold cross-validation with training parameters: batch size 2, patch size 128×128×128, initial learning rate 0.001, SGD optimizer, and 1000 rounds of training.

• The final best performance is still 3D UNet

Organization moduleinference process

• Inference is divided into two steps: organ segmentation first, and then soft tissue segmentation.

• The original image is downsampled for organ segmentation, and the result is used to generate "non-organ regions".

• The non-organ image is cropped into small pieces and input into the tissue network, and finally the whole structure is restored by splicing.

Whole body inference of UCHL1-EGFP mice | Quantification of neurological analysis of whole mice.

• Using a sliding window inference method and the ZARR file format, combined with the DASK parallel computing framework, the mouse whole body scan data was efficiently processed to ensure rapid analysis.

• Prior to Inference, the researchers performed percentile normalization on each scan image by calculating the 0.10th percentile and 99.9th percentile to set the minimum and maximum thresholds.

• The inferred whole-body neural segmentation results were analyzed for connectivity components, removing large false-positive regions caused by high-intensity regions (e.g., certain sites within the body of the mouse), the

• The researchers performed quantitative analyses of neural voxels and densities from three perspectives: whole body, individual tissues, and specific organs. To calculate whole-body neurovoxels, they combined the results of organ and tissue segmentation in the tissue segmentation module to create a binary mask that covered the entire mouse body. Neurovoxels and densities were then calculated by expanding this mask.

• Quantifying head nerves, the researchers created an accurate head nerve mask by overlaying the brain mask with the whole body mask, which allowed them to accurately calculate neural voxels within the head region.

Whole body inference in CD68-EGFP mice | Quantification of immune cell analysis in whole mice

• Regions of autofluorescence and CD68 channels were cropped from LSFM images using the same patch size and distribution as for tissue segmentation, and only patches located within the mouse were retained for subsequent analysis.

• A CD68+ segmentation network was used on the cropped patches to obtain a binary mask. connectivity component analysis was performed via the cc3d library to extract the positional information, volume, center of mass, and shape of each detected cluster. Clusters were categorized into internal organs or tissues based on center of mass location. Clusters that were not within these regions were removed, as well as those with overly elongated shapes (which may be vascular or neural artifacts).

• Next, clusters were categorized by volume. For each mouse and its organs, the percentage of clusters in different volume classes was analyzed, comparing the difference between the Chow and HFD groups.The CD68 segmentation network was still able to reason efficiently on certain new tissues that had not been seen before, showing zero sample migration.

• To validate the results, visual inspection and manual inspection using VR annotation were performed and compared with the automated segmentation results, and only data with DICE scores greater than 65% were included in the analysis.

Materials and steps of biochemical experiments, histological analyses and raw letter methods

🍎 Results.

Visualizing the nervous and immune systems in obesity

MouseMapper Whole Body Analysis

• The NerveModule demonstrated optimal segmentation performance by annotating the nerves of UCHL1-EGFP mice in Virtual Reality (VR) and using a 3D UNet deep learning model for training and testing. The Dice coefficient of the final model was 0.7913 ± 0.1423, demonstrating better neural segmentation.

• Immune-Module obtained better performance than other deep learning networks by labeling CD68-EGFP+ immune cells and training with the 3D UNet model. The model demonstrated zero-sample inference capability to successfully segment immune cells in unseen tissues such as liver and intestine.

• Tissue-Module is used for organ and tissue segmentation to improve computational efficiency by optimizing the training data, including the use of downsampled images. By labeling 20 organs in mice, the model performs well, especially in the comprehensive segmentation of organs. For segmentation of fat and muscle tissues, the use of full-resolution data is required to accurately detect tissue texture.

• By integrating these three modules, MouseMapper was able to generate a comprehensive whole-body anatomical map, revealing that obese mice have significantly increased adipose tissue (both visceral and subcutaneous fat) and liver volumes, as well as increased lymph node mass. This provides a unified frame of reference for quantitative analysis across regions.

Structural Changes Accompanying Obesity-Related Behavioral Deficits in the Facial Nerve | Changes in Peripheral Nerves

Structural changes in the infraorbital nerve and proteomic changes in the trigeminal ganglion | Trigeminal Nerve Analysis

Revealing obesity-induced systemic inflammatory state | Immune cell changes

🐿️ Conclusion

🤗 Insights.

1. Different objects of study (e.g., nerves, cells, and tissues themselves are differentiated in this article), and a careful analysis of one category can produce a separate paragraph in RESULTS.

2. It is important to learn to build on the strengths and avoid the weaknesses, and to prioritize the logic and flow of the article; although some results look important, they have been placed in an appendix, perhaps intentionally out of interest in the reader's reading.

3. The workload of this article also seems to be very large, need to learn from the study, do a good job of organizing the project, early pen.

4. Need to study the project's Github repo, especially the reusable parts of it.

Citation