AI MRI 3D

disease recognition

SW Medical Project. Alexander Belousov • 24.10.2024

Overview

AI desease recognition

will be base with using

sequnce of AI models,

trained with 3D MRI/CT imaging data

Development Tools

Tools

●For desease recognition

software develpment will

be uses next tools:

PyTourch, Tensorflow from

Google, ONNX from

Microsoft and libraies

MONAI SSL, Auto3DSeg

from NVidea company.

MONAI

●MONAI SSL (Self-

Supervised Learning) and

Auto3DSeg are two

advanced tools within the

MONAI (Medical Open

Network for AI)

framework, designed to

enhance medical image

analysis.

MONAI SSL is NVidia product

MONAI targeting description

●MONAI SSL focuses on self-supervised learning techniques,

which allow models to learn from unlabeled data. This is

particularly useful in medical imaging, where labeled data can be

scarce and expensive to obtain. By leveraging large amounts of

unlabeled data, MONAI SSL can improve the performance of

models on downstream tasks such as segmentation and

classification.

Auto3DSeg

Auto3DSeg is a comprehensive solution for large-scale 3D medical image segmentation.

It automates the process of developing and deploying segmentation algorithms, making it

accessible for both beginners and advanced researchers. Key features include:

Unified Framework: A self-contained solution requiring minimal user input. Flexible

Modular Design: Components can be used independently to meet different user needs.

Support for Custom Algorithms: Users can integrate their own algorithms into the

framework. High Accuracy and Efficiency: Achieves state-of-the-art performance in

various applications. Auto3DSeg has been tested on large-scale 3D medical imaging

datasets and has demonstrated high performance in several challenges, such as the

MICCAI 2023.

Goals for using

this development

tool

Applying MONAI SSL and

Auto3DSeg to cancer detection

in MRI scans can significantly

enhance the accuracy and

efficiency of identifying and

diagnosing cancerous tissues.

MONAI SSL for Cancer Detection

Feature Learning: Self-supervised learning models can learn intricate features from large datasets of unlabeled MRI scans.

This helps in identifying subtle patterns that may indicate the presence of cancer.

Pre-training Models: These models can be pre-trained on vast amounts of unlabeled data and then fine-tuned with labeled

data to improve performance on specific cancer detection tasks.

Auto3DSeg for Cancer Detection

Automated Tumor Segmentation: Auto3DSeg can automatically segment tumors from MRI scans, providing precise

boundaries of cancerous tissues. This is crucial for accurate diagnosis, treatment planning, and monitoring the progression

of cancer.

Multi-organ Segmentation: It can segment multiple organs and tissues simultaneously, which is useful in detecting

metastasis (spread of cancer to other parts of the body).

Customizable Pipelines: Researchers can customize the segmentation pipelines to focus on specific types of cancer, such as

brain tumors, breast cancer, or prostate cancer.

Practical Applications

Brain Cancer: Segmenting brain tumors to assist in surgical planning and radiation therapy.

Breast Cancer: Detecting and segmenting breast tumors to guide biopsy and treatment decisions.

Prostate Cancer: Identifying prostate tumors and assessing their extent to plan appropriate interventions.

Benefits

Improved Accuracy: Enhanced detection and segmentation accuracy lead to better diagnosis and treatment outcomes.

Efficiency: Automating the segmentation process reduces the time radiologists spend on manual annotation, allowing them

to focus on more complex cases.

Consistency: Automated tools provide consistent results, reducing variability in diagnosis and treatment planning.

These tools can transform cancer detection by providing more accurate, efficient, and consistent analysis of MRI scans,

ultimately improving patient outcomes.

We want to start with the following:

Urological Cancer Detection

Pre-training on Unlabeled Data: Models can be pre-trained on large datasets of urological

MRI scans without labels. This helps the model learn the underlying structure and features of

urological tissues.

Fine-tuning with Labeled Data: Once pre-trained, the model can be fine-tuned with labeled

data to specifically identify and classify urological tumors, improving accuracy and

robustness.

Automated Tumor Segmentation: Auto3DSeg can automatically segment urological tumors

from MRI scans, providing precise boundaries and volumes of the tumors. This is crucial for

planning biopsies, surgeries, and radiation therapy.

Multi-modal Integration: It can integrate different MRI modalities (e.g., T1, T2, DWI) to

improve the accuracy of tumor detection and segmentation.

Custom Pipelines: Researchers can customize the segmentation pipelines to focus on specific

types of urological tumors, such as prostate cancer.

Case Studies

Study 1: Prostate Cancer Segmentation

In a study using Auto3DSeg, researchers achieved high accuracy in segmenting prostate tumors from MRI scans. The automated tool provided consistent and

precise tumor boundaries, which were used to plan surgical resections and monitor treatment response.

Study 2: Multi-modal MRI Analysis

Another study utilized MONAI SSL to pre-train models on multi-modal MRI data, including T1, T2 sequences. The fine-tuned models demonstrated improved

performance in detecting and classifying prostate tumors compared to traditional supervised learning methods.

Multi-modal MRI data involves using different MRI sequences to capture various aspects of tissue properties, providing a comprehensive view of the area being

examined. Here are the key sequences:

T1-Weighted Imaging (T1)

T1 without Contrast: Provides detailed anatomical information. Tissues with high fat content (like white matter) appear bright, while fluids (like cerebrospinal

fluid) appear dark.

T1 with Gadolinium Contrast (T1Gd): Enhances the visibility of blood vessels and areas with a disrupted blood-brain barrier, such as tumors, making them

appear brighter.

T2-Weighted Imaging (T2)

T2 Imaging: Highlights differences in water content. Fluids appear bright, making it useful for detecting edema, inflammation, and other fluid-related

abnormalities.

Practical Applications

1st Stage:

Organs segmentation.

Get Prostate localisation image

2st Stage:

Classifies image

Make measurement of artifacts in image

Benefits

Enhanced Accuracy: Improved detection and segmentation accuracy lead to better diagnosis and treatment planning.

Efficiency: Automation reduces the time required for manual annotation, allowing radiologists to focus on more complex cases.

Consistency: Automated tools provide consistent results, reducing variability in diagnosis and treatment planning.

These advancements in urological cancer detection using MONAI SSL and Auto3DSeg are transforming the field, making it possible to diagnose and treat

urological tumors more effectively.

01.01.XX

Prepare SW

development tool

01.02.XX

Develop custom SW

code

15.05.XX

Train models on open

Datasets

01.06.XX

Train models on

anonymized Datasets

from hospital

01.10.XX

End user SW and work

with clients

Development schedule

About Us

Alexander Belousov - experienced software developer, (+972) 546-45-37-39, hodyrev@gmail.com

Experience of independent work. Extensive experience in the development of medical equipment. Participation in

projects for companies GE HealthCare, Philips Healthcare, Meta | Facebook

Alexander Belousov | LinkedIn

Sergey Klinov - experienced software developer, (+972) 542-07-16-16, sergklinov@gmail.com

Has experience in managing groups and departments in software companies

Sergey Klinov | LinkedIn