Automatic Detection and Motion Analysis of Cells in Body Fluids in the Diagnosis of Disease Using �Image Processing and Deep Learning

Jamiu Oluwaseun Ojeleye�359704

Supervisor:�Prof. Dr. Murat Ekinci

Jury Members: �Prof. Dr. Murat Ekinci, Doç. Dr. Güzin Ulutaş and Dr. Öğrt. Üyesi Selen Ayas

Problem Statement

Traditional way of counting and tracking blood cells in body fluid for the diagnosis of disease is labor intensive, time consuming and lacks precision and reliability.

A desktop based application that automates the method of counting and tracking of blood cells would be developed using image processing and deep learning techniques.

Solution

Solution

Design Overview

Images or video frames are accepted as input and they are processed by three major subsystems in the application

Subsystem Overview

1. Hemocytometer Grids Detection System:

Hemocytometer Grid Detection System

Hemocytometer Grid Detection System

Adaptive Canny Edge Detection

Hemocytometer Grid Detection System

Line Detection with Hough Transform

Hemocytometer Grid Detection System

Detection of Grids

Subsystem Overview

2. Cells Detection System:

Steps:�1. Dataset Gathering and Labelling using an implemented Annotation Tool.�2. Training deeplearning model using Darknet Library.�3. Implementing the cells detection system using a well optimized version of Darknet library.

Annotation Tool:

An Annotation tool was implemented in Visual Studio C++ for labeling dataset using the YOLO Format and “AT” Format

Cells Detection System

Training Data and Data Augmentation:

Cells Detection System

Note: Why are images splitted into two overlapping images?

• Training Large image with small objects requires large network size which requires alot of computing resources (GPU memory) and computational time.
• Resizing the large image in order to train with small network size results in loss of smaller objects details and accuracy of model.

​

​

Cell Detector Network Architectures:

Cells Detection System

YOLOv3 Architecture:

Seun-1 Architecture:

Cells Detector: � CSF Frame + Deeplearning Model + Optimized Darknet Library

Cells Detection System

Steps:

• Split image into two overlapping images of size 1080 x 1080
• Detect cells in the two Images using YOLOv3 or Seun-1
• Merge boundboxes of detect cells in the two images and remove overlapping boundboxes
• Reconstruct boundboxes of original image from the resulting boundboxes from the previous step

3. Cells Motion Analysis Systems:

Subsystem Overview

1. Cells Counting System

2. Cells Tracking System

Cells Counting System: � Hemocytometer Grid Detection + Cells Detection System

Cells Motion Analysis System

Cells Motion Analysis System

Cells Tracker: � Cells Detection System � +� Tracking Algorithm

Cells Tracking Algorithm: Mean Shift Tracking + Normalized Cross-Correlation(NCC)

Cells Motion Analysis System

Steps:

• Check if current frame and next frame are similar using Normalized Cross-Correlation(NCC) Algorithm.
• If they are similar, track cells in next frame using Mean-Shift Algorithm. Otherwise, re-detect the cells in next frame using the deeplearning model.

Mean Shift Tracking Algorithm for tracking a cell with position Y0 in Frame t

Cells Motion Analysis System

Note: The edge information of the images were incorporated to the joint-color histogram

Illustration:

Real-time Face Tracking using Mean-Shift Algorithm

Cells Motion Analysis System

Cells Tracking System:

Cells Motion Analysis System

Demo

Demo Video: https://drive.google.com/file/d/1DkLVzRYsF0BvHsSgTYXInB_TnhB7X4ql/view?usp=sharing

Thank You For Your Time