A NOVEL OBJECT DETECTION MODEL

(YOLOv5) FOR IMPROVED LUNG NODULE

IDENTIFICATION IN MEDICAL IMAGES
G.Eswari1,A.Thangamani2, S. Gopinath3,R.Harish4
1,2,3,4
Department of Information Technology, Velammal Institute of Technology, Panchetti, Chennai, India.
1
geswari.itvelammalitech.edu.in, [email protected], [email protected],
4
[email protected]

Abstract- This paper introduces a novel approach In recent years, deep learning techniques have
utilizing the YOLOv5 object detection model for shown promising results in various medical image
enhanced lung nodule identification in medical analysis tasks, including lung nodule detection.
images. Lung nodules are early signs of lung YOLO (You Only Look Once) is a popular object
cancer, and their accurate detection is crucial for detection algorithm known for its speed and
timely diagnosis and treatment. Traditional accuracy. In this paper, we propose the utilization
methods for nodule detection often rely on manual of YOLOv5, an upgraded version of the YOLO
interpretation or computationally expensive algorithm, for the detection of lung nodules in CT
algorithms. In contrast, the proposed YOLOv5- scans. The primary motivation behind this research
based model offers real-time detection with high is to develop a robust and efficient system that can
accuracy, thus potentially improving patient accurately identify lung nodules in medical images,
outcomes. Experimental results demonstrate the thereby assisting radiologists in their diagnostic
effectiveness of the proposed model in accurately process. By leveraging the advancements in deep
identifying lung nodules from medical images, learning and object detection, we aim to overcome
showcasing its potential as a valuable tool for the limitations of traditional methods and provide a
radiologists and clinicians. In contrast, the reliable tool for early lung cancer detection. The
proposed YOLOv5-CNN model offers real-time remainder of this paper is organized as follows:
detection and classification with high accuracy, Section II provides a review of related work in lung
potentially improving patient outcomes. nodule detection. Section III details the
Experimental results demonstrate the model's methodology, including dataset preparation and
effectiveness in accurately identifying and model architecture. Section IV presents the
classifying lung nodules from medical images, experimental results and performance evaluation of
underscoring its promise as a valuable tool for the proposed model. Finally, Section V concludes
radiologists and clinicians. the paper with a discussion on future research
directions. The suggested YOLOv5-CNN model,
on the other hand, provides high accuracy real-time
Keywords: Data Set (medical images), detection and classification, which may enhance
YOLOv5(You Only Look Once), Object Detection, patient outcomes. The model shows potential as a
Python, GLCM method, Convolutional Neural useful tool for radiologists and physicians since
Network (CNN), Lung Nodule Identification. experimental findings show how well it can
recognize and categorize lung nodules from
medical images.
1. INTRODUCTION
2.LITERATURE SURVEY
Lung cancer remains one of the leading causes of
cancer-related deaths worldwide, with early [1] Xia, X., & Zhang, R. (2023). “A novel lung
detection playing a pivotal role in improving nodule accurate segmentation of PET-CT images
patient survival rates. Computed Tomography (CT) based on Convolutional neural network and Graph
scans are commonly used for lung cancer Model”, PET/CT (positron emission
screening, as they provide detailed images of the tomography/computed tomography) imaging can
chest area. However, the manual interpretation of provide information on the functional features of
these scans is time-consuming and prone to human metabolism as well as the physical location of the
error. Automated methods for lung nodule patient's body. In this work, we have proposed a
detection have thus garnered significant interest in novel deep learning-based graph model network
the medical imaging community.
which can automatically fuse dual-modality the lung nodule using a deep learning U-net
information for tumor area segmentation. This technique. A 3D U-net that was trained with a
process assumes that distinct scales are acquired novel loss function to address under- or over-
by the spatially variable features, which encode segmentation problems provides even more
different feature information over different improvement. As shown by the Dice coefficient
modalities. The outcomes demonstrated how well (DCE), the resulting segmentation accurately
our approach combines data from the PET and CT depicts lung nodules in terms of form and volume
modalities, with segmentation accuracy results in metrics.
DSC of 0.86 and sensitivity findings of 0.83—a
3.61% improvement over the W-Net. [5] Fan, X., Lu, Y., Hou, J., Lin, F., Huang, Q.,
& Yan, C. “DMC-UNet-based segmentation of
[2] Saihood, A., Karshenas, H., &Naghsh-Nilchi, lung nodules”. An enhanced network built on
A.R. “Multi-Orientation Local Texture Features UNet is the DMC-UNet network that is suggested
for Guided Attention-Based Fusion in Lung in this study. The DMC-UNet X/Y Channel
Nodule Classification”. This study offers a great
Attention Module, Coordinate Attention (CCA)
degree of flexibility in focusing on pertinent data
attention mechanism, lightweight residual
that is retrieved from various nodule regions in a
structure, and multiscale feature up sampling
non-local way. Additionally, in order to give the
fusion are all combined in this network. The
model more precise discriminative data from the
general structure of the order to reduce the number
nodule volume, specially constructed local
of parameters and computation required for the
Multiple orientations of texture feature descriptors
(TFDs) are derived from the nodule slices. model and improve the efficiency of both model
training and prediction, the network first replaces
[3] Song, J., Huang, S. C., Kelly, B., Liao, G., the convolutional units of U-Net with residual
Shi, J., Wu, N., ... & Yeom, K. W.” Automatic units. The incorporation of a CCA attention
lung nodule segmentation and Intra-Nodular mechanism following the up-sampling fusion can
Heterogeneity Image Generation”. In this work, improve the recovery of spatial information. l
we offer an end-to-end architecture to produce Convolutional Neural (ESPCN) model, which
intra-nodular heterogeneity and perform fully seeks to broaden the perceptual field.
automated segmentation of various types of lung
nodules pictures intended for medical use. To do [6] Sivakumar, A., Suganthy, M.,
this, a Faster R-CNN model based on generalized Gunasundari, B., Sivamurugan, S., & Sajiv, G.
intersection over union loss in generative "A Robust and Novel Hybrid Deep Learning
adversarial network is introduced to examine a based Lung Nodule Identification on CT Scan
hybrid loss. For this investigation, 3,200 lung Images”. The aim of this study was to develop and
nodules from five hospitals were integrated with evaluate an image-based feature-based Novel
2,635 lung nodules from the Lung Image Database Hybrid Neural Network (NHNN) model for
Consortium image collection dataset. On the test nodule detection. This gives the radiologist
dataset, the suggested model produced an average additional information that can be utilized to find
dice coefficient (DC) of 82.05% when compared cancer when it is still contained within the nodule,
to manual segmentation by radiologists. in the early stages of the disease. Radiologists will
compare the CT scan data side by side to ascertain
[4] Youssef, B., Alksas, A., Shalaby, A., the location and condition of the nodule. To detect
Mahmoud, A., Van Bogaert, E., Alghamdi, N. S., the tumor in CT scan images and provide doctors
& El-Baz,A. “Integrated deep learning and with a comprehensible computerized assessment
stochastic models for accurate segmentation of report, a dependable technique is needed.
lung nodules from computed tomography images:
a novel framework”. To accurately extract lung [7] Zhang, J., & Chung, T. M. “An Improved
nodules from 3D computed tomography (CT) YOLO V5 Model for Pulmonary Nodule
scans, a novel approach is presented in this study. Detection with Synthetic Data Generated by
Our method includes two crucial preprocessing GAN”. Finding pulmonary nodules is essential to
steps to improve nodule segmentation and search making the diagnosis of lung cancer since they
accuracy. First, we utilize a two-level joint are an early sign of the disease. Although a lot
Markov-Gibbs random field (MGRF) model to of study has been done and some progress has
been achieved, pulmonary nodules remain
identify the lung area, successfully differentiating
challenging to detect due to their hazy
lung wall nodules from the chest area based on
boundaries, tiny size, and variety of shapes.
similar visual features. Next, we minimize the Furthermore, medical databases are valuable
inclusion of neighboring lung tissues by and hard to come by. We provide an enhanced
identifying the region of interest (ROI) containing YOLO V5 framework for pulmonary nodule
diagnosis based on synthetic data from many
generative adversarial networks (GANs) in order
to address the aforementioned issues. Our
approach involves utilizing Resnet50 as the
YOLO V5 backbone. This eliminates the
gradient explosion issue and adds residual
connections for the purpose of retrieving
pulmonary nodule features.