Project Name: Center of Excellence in Artificial Intelligence for Medical Image Segmentation

Proforma for submission of proposal under the scheme of research and development

Ministry of Science & Technology Department of Science & Technology

“Fund for Improvement of S&T Infrastructure in Universities and Higher Educational Institutions (FIST) Program – 2021”

Project Abstract

In an interdisciplinary team, we will develop machine learning models to address problems in medical data analysis and prediction.
This will involve the acquisition, development, and application of machine learning techniques. Further, through this experience, students
will be able to gain a solid understanding of deep learning models (aka neural networks) as well as their practical applications to medical
imaging data.
A biomedical image segmentation technology is an emerging biomedical image processing technology that has contributed greatly to
sustainable medical care. It is now becoming an important research topic in computer vision. Computer vision has become advanced due to
the development of deep networks that have solved problems poorly handled by classical image processing techniques. In general, we can
claim that Deep Neural Networks (DNN) are superior in image recognition tasks, which include object recognition, image classification,
segmentation, and activity recognition. In this work, deep neural networks meet medical imaging to create an engaging interface. Our research
will focus on magnetic resonance image segmentation, microscopic medical images, which is one of the most popular tasks because it offers
the most well-structured datasets to be analyzed. In order to diagnose, monitor, and plan treatments, medical imaging (MRI) will be utilized.
Manual practices require anatomical knowledge and are costly and time-consuming. Detecting and diagnosing leukemia as early as possible,
that is, distinguished by a sensitive and accurate test, with minimum cost, presents a major challenge in the field of disease diagnosis. We
conduct this systematic review to identify studies using Machine Learning and Deep Learning techniques for the analysis of peripheral blood
smear (PBS) images in order to detect leukemia. Segmentation of images is a common technique in research and medical imaging. The goal
of segmentation is to obtain a visual representation of the blast in terms of the structure and its features. Techniques for segmentation include
thresholding, boundary-based segmentation, region-based segmentation, and hybrid techniques that combine boundary and region principles,
and most methods combine boundary and region criteria. Image analysis of blood smears is crucial in diagnosing various blood diseases. The
Deep Learning technique determines the origin of a blast by analyzing its sequence. Furthermore, they are inaccurate due to human error.
The issues associated with biomedical imaging to help clinicians, radiologists, pathologists, and clinical researchers diagnose and understand
various diseases affecting people. The rapid development of artificial intelligence, specifically deep learning, has benefited machine learning
and computer vision algorithms in the field of image segmentation. Segmenting medical images using deep learning can help doctors
accurately determine the size of tumors, quantitatively evaluate the effects of treatment and significantly reduce doctors' workload.

Keywords: Image segmentation; Deep learning; Convolutional Neural Network; Medical image

N.M.A.M INSTITTUTE OF TECHNOLOGY NITTE, KARKALA, KARNATAKA INDIA 1

1) N.M.A.M Institute of Technology Nitte, Karkala, Visvesvaraya Technological University, Karnataka, INDIA
2) K.S. Hegde Medical Academy, Mangalore, Nitte Deemed to be University, Karnataka, INDIA

Three modules have been developed for the project

Module 1: Image Sources

Objective:

A primary objective of image enhancement is to make the original image more suitable for a specific application than the enhanced image.
Digital image enhancement techniques offer a multitude of ways to improve the visual quality of images. Enhancing an image involves
improving the way information is communicated in images for human viewers or providing better input for other automated image processing
techniques. Methods with an on-pixel domain, such as methods using the Fourier transform of an image, and those with a frequency domain.
In digital image processing, the first step involves digitizing an image and converting it to a computer image file that can be stored in a
computer's memory or on a storage medium, like a hard drive. It is done by converting the image into a numerical code that a computer can
understand. This can be accomplished by scanning or linking a video camera to a frame grabber board in the computer.

Scope of work to be carried out:

Camera Technology: A comparison of CCD resolution, signal strength, pixel array, color capturing, and multiple passes over 3-CCD
systems.

Scanners: Magnetic Resonance Image (MRI), Computed Tomography (CT), Ultrasound, Microscopic Images.

Synthetic Images: Images and formats represented by computers: Color representation schemes: RGB, YCrCb, YUV RGB, JPEG, TIFF.

Image manipulation operations

1) Fundamental Steps: Spatial frequency domain operation, point-based operation, and kernel-based operation
2) Enhancing Image Quality: The best way to improve your images is to denoise, deblur, and equalize the histograms. Image
denoising with a denoising autoencoder.
3) Geometric transformations: Affine transformation in 2D or 3D, and homogeneous coordinates. Applying perspective
transformation and homography. Transforming color space (RGB → Lab)
4) Image Analysis: A segmentation algorithm (pixel values, watershed algorithm, a calorimetric algorithm, deep neural network
algorithm), pattern recognition (pattern matching, neural network-based recognition). Deep semantic segmentation, Deep instance
segmentation, Image segmentation with self-organizing maps.

N.M.A.M INSTITTUTE OF TECHNOLOGY NITTE, KARKALA, KARNATAKA INDIA 2

Objectives:

The purpose of image segmentation is to extract the part of an image which is of interest & to subdivide it into constituent parts. Image
segmentation is part of an image analysis system. The process of segmenting an image is a difficult one that is affected by several aspects,
such as noise, low contrast, and irregular boundaries of the objects. Segmentation of an image allows the identification of objects and
boundaries, which are typically found at the edges of a picture. The concept of segmenting images in medical imaging is the process of
analyzing and processing images using computer image processing technology to segment, extract, reconstruct, and display images of human
organs, tissues, and diseases in three dimensions. In this method, images are divided into several groups based on similarity or difference
between them, and medical practitioners can analyze lesions and other areas of interest in a qualitative or quantitative way by using this
method. Through image segmentation, feature measurement, and object representation, image segmentation seeks to extract information
within an image that is expressed as data.

Scope of work to be carried out:

1) When using machine learning to process medical images, the data set is often divided into three parts: training set, validation set,

and testing set. The training set is used to train the network model, the verification set to adjust the hyperparameters of the model,

and the test set to verify the final effect of the model.

2) In addition to histograms, other segmentation schemes based on pixel progression are known to be extremely well-organized for

evaluating them as only one exceed in each pixel is required.

3) To expand the image data set, preprocessing the image, which generally including standardization of the input image, random

rotations, and random scaling.

4) To measure and evaluate the effectiveness of medical image segmentation, performance indicators must be established and verified.

This is an essential component of the process.

5) Building blocks of Convolution Neural Network: Convolutional features are modeled in a way that models the channel dependence,

so this structure should enhance representation ability.

N.M.A.M INSTITTUTE OF TECHNOLOGY NITTE, KARKALA, KARNATAKA INDIA 3

Objective:

Optimization functions for classification and regression problems are divided based on the deep learning model's performance and are
reduced to improve the performance. Shallow Neural Network is composed of only one hidden layer, input and output layers, whereas Deep
L-Layered Neural Network could have L number of small hidden layers. As we are working on automatic brain MRI segmentation, it is
important to briefly describe the basic brain structures among which are trying to distinguish a) White matter (WM), b) Grey matter (GM),
c) Cerebrospinal fluid (CSF). The main System of anatomical coordinates: the axial plane, sagittal plane, and coronal plane. Working on
image synthesis on MRI images is the process of applying intensity transformation to a given set of input images to produce new biopsies or
tissue contrast.

Scope of work to be carried out:

1) Content-based image retrieval (CBIR) to assist them when making decisions about images.
2) The focus of deep learning in MRI for segmentation and classification of reconstructed magnitude images.
3) The use of dropout during inference is a powerful technique for improving model performance in CNN
4) Training the data by combining the hand-engineered features with CNN features is a very reasonable approach. When a large
amount of training data is lacking, it is difficult to learn good features automatically.
5) In leukemia diagnosis or detection, segmentation is critically important. The reliability of feature extraction and the classification of
leukemias are proportionately affected by the accurate segmentation of the lymphocytes.
6) The use of image augmentation for leukemia detection may lead to better coverage of the data space and, thus, improved results
when using these methods.
7) Future research can be directed toward using ML algorithms, particularly deep-learning algorithms, in computer-aided detection
(CAD) systems, whole-slide imaging (WSI), and even apps and software to assist the pathologists and oncologists working at
hematology laboratories.
8) Segmenting semantic images involves assigning a class to every pixel of an image. In other words, dense prediction describes this
process since each pixel is assigned a class. Analyses performed by radiologists can be enhanced by machines, thereby reducing the
time needed to do diagnostic tests.
9) Applying U-net architecture for image segmentation in biomedical applications. The semantic segmentation to be successful, it is
necessary to not only discriminate pixels, but also project the discriminative features learned at the different stages of the encoder
onto pixels.

N.M.A.M INSTITTUTE OF TECHNOLOGY NITTE, KARKALA, KARNATAKA INDIA 4

Module 1: Image enhancement

Digital image enhancement involves adjusting images so that they are suitable for appearance or further analysis, for example, removing noise
or brightening images to make them easier to analyze or display. A technique called histogram equalization is widely used to improve the
contrast in images. It works by spreading out the most frequent values of intensity, i.e., spreading out the intensity range of an image.
The following methods of morphological filtering are used:
• Histogram equalization
• Wiener filter noise removal
• Median filtering
• Uniform mask filtering
• Contrast-limited adaptive histogram equalization (CLAHE)
• Decorrelations stretch
Filtering and enhancing techniques can be divided into several categories based on their overall functions. Among these functions are
detecting, extracting, or separating signals, reducing noise, and emphasizing certain aspects of a signal.
• Point/histogram operations
• Time/spatial domain operations
• Frequency domain operations
• Geometric operations

Module 2: Medical Image Segmentation

The objective of image segmentation is to simplify or alter the representation of an image so that it can be analyzed more clearly and effectively.
It's typically used to find objects and boundaries (lines, curves, etc.) in an image. A segmented image includes segments of different objects
or parts of objects, each pixel assigned to one of these categories. The use of image segmentation techniques makes this problem easier for
us to solve by providing us with more details.

• A group of pixels in the same category will share a similar set of greyscale values and form a connected region.
• A neighboring pixel in a different category has a different value from its neighbor.

Segmentation can be divided into three categories: thresholding, edge-based methods, and region-based methods
• The thresholding process involves allocating pixels to categories based on their values.
• A filter called an edge filter is applied to the image, pixels are classified according to the output of the filter, and pixels which are
not separated by an edge are grouped together.
Advantages: It is good for images having better contrast between objects.
• In region-based segmentation, pixels that are neighbor’s and have similar values are grouped together iteratively, and pixels with
dissimilar values are split up
Advantages: Simple Calculations and Fast operation
• A segmentation technique based on clustering divides pixels in an image into clusters of homogeneous pixels
Advantages: This program generates excellent clusters when using small datasets.
• Masking R-CNN outputs three information about each object in the image: the class, bounding box coordinates, and the object
mask
Advantages: An easy, flexible, and general method of image segmentation. It is also the current state-of-the-art.

N.M.A.M INSTITTUTE OF TECHNOLOGY NITTE, KARKALA, KARNATAKA INDIA 5

The deep learning algorithm passes data through several layers of neural network algorithms, each of which provides a simplified
representation of the data to the next.

• Deep Learning performs better with large data, but with smaller data, traditional machine learning algorithms perform better.
• The benefits of Deep Learning techniques outweigh other methods if there is ignorance of the domain for feature introspection.
• A deep learning algorithm is especially useful when it comes to solving complex problems such as image classification, natural
language processing, and speech recognition.

Object Detection: This is a variation of the image classification in which bounding boxes are used to illustrate each object's position. The
objective of object detection is to classify the image with localization and to detect objects within it, even if the image contains multiple
objects with localization and classification.
• Challenging: Determining the boundaries of an infected area and labeling its contents
Object Semantic Segmentation: The process of identifying pixels assigned to each object within an image rather than drawing boxes
around each one as in object detection.
• Challenging: Identifying objects (instances) in groups (each object in a group) belonging to the same class.
Image Classification with Localization: The basic idea of localization in image classification is to assign a class label to the image and
outline the outline of the object in the image by drawing a box around it (showing it in a reference frame).
• Challenging: Marking an x-ray as cancerous or not and drawing a box around the infected area.
Image Colorization: The process of neural colorization involves converting grayscale images into full-color images.
• Challenging: A task like this can be compared to an objectively evaluable filter or transformation.
Image Reconstruction: Filling in missing or corrupt parts of an image is part of image reconstruction and image inpainting
• Challenging: This is a task that is like a filter or transform in a photograph that probably does not have an objective evaluation

N.M.A.M INSTITTUTE OF TECHNOLOGY NITTE, KARKALA, KARNATAKA INDIA 6