Segmentation of Body fat using Machine Learning

Project Title

Segmentation of Body fat using Machine Learning

Project Area of Specialization Artificial IntelligenceProject Summary

Obesity is increasingly prevalent and associated with increased risk of developing type 2 diabetes, cardiovascular diseases, and cancer. Magnetic resonance imaging (MRI) is an accurate method for determination of body fat volume and distribution. However, quantifying body fat from numerous MRI slices is tedious and time-consuming. Here we are developing a machine/deep learning-based method for measuring visceral and subcutaneous fat.

In first phase, we convert CT (Computed Tomography) images into MRI (Magnetic resonance imaging) images, as it is the gold standard in medical imaging. Unfortunately, it is not a viable option for patients with metal implants, as the metal in the machine could interfere with the results and the patient’s safety. So, to solve this problem we are using CT images and then converting in MRI images.

In second phase, we are using segmentation techniques to segment out body fat. A machine/deep learning approach was developed for segmenting visceral and subcutaneous fat based on the U-net architecture. Machine/deep learning can learn patterns in visual inputs in order to predict object classes that make up an image.

Project Objectives

The main aim of this project is to design a model for Fat segmentation. The objective set are:

• Detection of fat region in Human Body.
• Separation of Brown Adipose Tissues (BAT) and White Adipose Tissues (WAT).
• Developing and analyzing Machine Learning Algorithms for Fat segmentation
• To run and verify a model with dataset having no ground truth i.e. images or results from experts.

Project Implementation Method

Our objective is to detect the human body fat, segment brown and white fat using Machine Learning Techniques. For this purpose, we need dataset of MRI Scanned Images, which we will acquire by translating labeled dataset of CT Scanned Images. The prior step before the translation is the pre-processing to remove the noise and normalized the MRI images. Image Translation is achieve using CycleGAN tool. Segmentation part is done by using the U-NET which is neural network for segmentation of biomedical images.

For training purpose, the available dataset is CT images. CT is preferable for bones and MRI is used for soft tissues such as fat parts around the organs. Due to unavailability of MRI dataset, we need to translate these images. CycleGAN uses unsupervised method using the training dataset and move from source to target domain.

U-NET is used for semantic segmentation, which does label the ROI in the image to equivalent class. U-NET model works on Convolution, Max Pooling and Transposed Convolution.

Software Techniques used in project Serial # Components Description  1. CycleGAN ML Technique for Image Translation 2.  U-NET Technique for Image Segmentation 3.  ITK-Snap/3D-Slicer Tool for Segmentation of CT/MRI images (for Verifications of results)

The framework of our project is as mentioned in above table, we will use CycleGan for Image Translation from one domain to another domain. And then we will use these translated images for segmentation in U-net. We are using deep learning techniques so we will not use ITK-Snap as much in our project. But we will use ITK-Snap to compare our generated results of U-net with the segmented image which is produced by using ITK-Snap.

Serial #1.2. 3. Benefits of the Project

1. Our project is beneficial in medical perspectives i.e. to prevent from diseases such as:

• Obesity
• Cardiac diseases
• Diabetes
• Some Metabolic syndromes

2. Useful for datasets having no ground truth, which means we can train our model using the training data of one domain and then validate that model by using the data from other domain.

3. All existing algorithms dependent upon expert knowledge, but our model can segment out fat using machine learning techniques with no need of any expert.

Technical Details of Final Deliverable

1. Our product will be a complete model which can segment out body fat tissues from MRI/CT images.
2. Firther it can be used for images (datasets) with no ground truth available.

Final Deliverable of the Project Software SystemCore Industry MedicalOther Industries Education Core Technology Artificial Intelligence(AI)Other TechnologiesSustainable Development Goals Good Health and Well-Being for People, Industry, Innovation and InfrastructureRequired Resources

Elapsed time in (days or weeks or month or quarter) since start of the project	Milestone	Deliverable
Month 1	Initial Study and Project Research	Report on initial study and existing models for fat segmentation
Month 2	Model Selection and Comparison	Comparison report for different Neural Networks for image segmentation and translation.
Month 3	Model Selection for Image translation	Selected Neural network for image translation i.e. CycleGAN technique
Month 4	Model Selection for Segmentation	Selected Neural Network for Segmentation i.e. U-net architecture
Month 5	Selected models analysis	Report of training with previous data, their results and accuracy for selected models.
Month 6	Training for our own Dataset	Trained U-net for our own data set
Month 7	Translation from CT to MRI images	Trained CycleGAN and translated images from CT to MRI
Month 8	Final model for fat Segmentation	Report on final outcome and results for verification.