Web Based AI

Solution For

Breast Ultrasound

Team

Members

Supervisors

Ahmed Mostafa

Saied Salem

Omar Mansour

Yasien Essam

Dr. Ahmed Ehab

Dr. Tamer Basha

Sponsored By

Motivation

• Increasing Numbers of new cases over the years

​

• Shortage of physician Numbers

• Subjectivity and Variability of doctor decision

​

• Time-consuming process for Assessment

​

• At early stages cancer survival rate is 99%

https://www.nationalbreastcancer.org/early-detection-of-breast-cancer/#:~:text=According%20to%20the%20American%20Cancer,clinical%20breast%20exams%20and%20mammograms

Number of new occurrences internationally

2020

Number of new occurrences

In Egypt

2020

Statistics

Reference: https://gco.iarc.fr/today/data/factsheets/populations/818-egypt-fact-sheets.pdf

https://www.hindawi.com/journals/bmri/2022/9605439/fig3/

BI-RADS

• BI-RADS scoring system assigns a numerical score ranging from 0 to 6

​

• Developed by American College of Radiology

​

• Each number corresponds to different levels of malignancy probability according to certain acoustic and geometrical features

BI-RAD 5

BI-RAD 2

BI-RAD 3

​

BI-RAD 4

​

Solutions

• Empower the workflow via AI

​

• Improve the accuracy of breast cancer detection

​

• Increase time-efficiency

​

• Provide a solution anywhere and anytime

Market �Research

02

Stakeholder Meeting

• Compliance with regulatory requirements and industry standards

​

• Using Segmentation Or Detection?

​

• Desktop or Web-Based?

Global Competitors

​

​

Local Competitors

​

​

Why Ultrasound?

No Ionizing Radiation

Portable, Mobile & Widely Used

Cost-Effective

Flexibility

Reference: Dense Breasts: Answers to Commonly Asked Questions - NCI (cancer.gov)

Why Web-Based?

01

• Easier To Receive New Improvements

• On Premise Data Center
• VPN

• Windows or MAC with different browsers , desktop, tablet, etc.

General Access To Medical Images

Security

Compatibility

Flexibility

• Hospital PACS
• Cloud Storage

02

03

04

Business Model

03

Cloud Infrastructure Costs

Powerful GPUs for Development

Costs

Resources

Business Model

​

System Specs

04

Scenario

Take Acquisition

PACS SYSTEM

App

Viewer

Log In

Operation

Specs

System Walkthrough

System Main Features

Sign Up

Account Verification

Login Page

Reset Password

ExamList

Content:

​

• Patient Information
• Malignancy Probability
• Classification
• Study Status

​

Features:

​

• Filter Functionality
• Global Search
• Exam Info Edit
• Pagination

Add Examination Feature

Upload PNG Locally

Browse from local directory to upload PNG. Application then convert it to DICOM

Upload DICOM Locally

Browse from local directory to upload DICOM Files

Saturn is the ringed planet. It’s composed mostly of hydrogen and helium

PACS Connection

Using Orthanc to simulate uploading from modality via PACS

Upload Multiple Files

User can upload multiple DICOM & PNG at the same time.

SECURITY FEATURES

• Email Verification : Sending a confirmation single-code via email, and requiring the user to write it to complete the account setup

​

​

• Session Timeout : If the user is inactive for 15 minutes, the user will be forced out to the login page.

​

1. Report Generation with Annotations, Physician Notes & Final Decision

​

1. Show & Hide AI Diagnosis

3. Lesions Table

Thumbnails displaying DICOM instants within the study

Annotations & Measurements Tools: Panning/Zooming/ROI (Ellipse-Rectangle-Free Hand-Line Angle)/Contrast/MultiFrame Support

AI Segmentation Contour

Doctor Measurements

Report Generation

Multi Frame Support

Technical

Specs

System Architecture

Cloud Performance

Cloud Processing Time

VM Specs:

• 4 vCPUs
• 32 GB RAM
• 50 GB SSD Disk

Client Sequence Diagram

Server-Side Operations

Database Management

• Design & Manage
• Optimize performance

​

Authentication & Authorization

• Verifying users (JWT)
• User's permissions to resources

​

APIs

• Way of communication between server and client
• Expose certain functionality

​

PACs Integration

• Integrating with Orthanc that is an open-source PACS through REST APIs

​

Server-Side Roles (Database)

• Admin will have the authority for everything.

​

• Doctor can view his own patient

​

• Patient has its studies.

​

• Studies has its own AI output as mask probability.

​

Tools

Client Side

Server Side

AI

PACS

Cloud

VM OS

Dockerization

​

​

• Using Docker :
• Portability
• Scalability
• Microservices Architecture

​

• Volume should be out of the container to maintain data

Application Performance Testing

Stress Testing

• Test server side using PyTest.
• Stress Testing Using Locust server
• The application is capable of handling up to 500 users.
• Request per second is 130 with 0 failure.

Cloud Deployment

​

• Using Microsoft Azure as a cloud server

​

• Using Ubuntu VM for deploying the application

​

• Using NGINX to solve CORS (Cross Origin Resource Sharing) problem by reverse proxy

​

Methodology

05

Literature Review

Enhanced Small Tumor-Aware Network for Breast Ultrasound Image Segmentation : ESTAN Architecture

Attention-Enriched Deep Learning Model for Breast Tumor Segmentation

​

in Ultrasound Images

​

Hover-Trans (Classification SOTA)

Datasets

Data Information

Total Data Distribution

3483 Benign

2098 Malignant

133 Normal

Challenges

Breast Ultrasound Images are known for :

​

1. High Variability & Less Uniform
2. Requires Experience
3. Full Of Noise
4. Black Border and UI Components
5. Highly Imbalanced Data

UI Removal Algorithm

AI Pipeline

AI Pipeline

Models

Classification Model

Classification Model

Model

Efficient-Net V2-B0

(270 Layers)

Original Data

1170 B-1531 M-103 N

4 Datasets

224x224x3 Shape

16 Batch Size

Learning Rate

Decaying Cyclic Learning Rate

Loss Function

Cross Entropy Or Focal Loss

Augmented Data

3399 Samples

​

Train-Val-Test

3399/551/586 Samples

Comparison Between Classifiers

Why Efficient-Net

Loss Functions

Cross-Entropy

Better With Balanced Dataset

Better With Imbalanced

Datasets

Focal Loss

Metrics On Test Data

179 Benign

392 Malignant

586 Samples

80%

Success Rate

89%

F1-Score & AUC

We achieved F1 Score of 0.89 and AUC of 0.97 on Test Data

Results

The classifier also reached an average specificity of 92 and test cross entropy of 0.269

Precision

0.89

Recall

0.89

15 Normal

Test Metrics

GradCams

Metrics On Astute Data

2370 Benign

401 Malignant

2771 Samples

80%

Success Rate

85%

F1-Score

We achieved F1 Score of 0.86 on Real Life Data

Results

The classifier can successfully generalize well

Precision

0.86

Recall

0.85

Astute Confusion Matrix

Segmentation Model

Segmentation Model

Model

Attention U-net

Input Data

5720 Samples

All Datasets

128x128x3 Shape

32 Batch Size

Learning Rate

Reduce On Plateau

Schedule

Loss Function

Log Cosh loss

Augmented Data

9148 Samples

Train-Val-Test

9148/573/573 Samples

Model Architecture

Segmentation Metrics

Intersection Over

Union

Dice Score

Loss Functions

Combo Loss

More flexible than dice and better highly imbalanced task

​

Overcome non convex nature of dice loss

Log Cosh Loss

Pytorch Data Engine

Initial Comparison Between U-Nets

Final Comparison Between U-Nets

Metrics On Test Data

573 Samples

80%

0.85

Dice Score

Results

Model also reached 0.15 dice loss on test data

Val Score

0.86

Test Score

0.85

Test Segmentation

Cloud Performance

Cloud Processing Time

VM Specs:

• 4 vCPUs
• 32 GB RAM
• 50 GB SSD Disk

Future Work

06

Future Work

• Publishing 2 papers (System & AI)

​

• Hospital Deployment

​

• Compatibility With HL7

​

• PACS Filtering

​

• Compete In Nile University Competition

Thank You