Design and Development of an AI Model to

Assist Pregnant Women on Goitre Nodules for

Diagnosis Using Convolutional Network - A
Deep Learning Approach
Prasad Vadamodul1 Jahnavi Jallepallit2 Aravind Malothu3
Department of Computer Science Department of Computer Science Department of Computer Science
& Engineering, & Engineering, & Engineering,
1,2 1,2 1,2
GMR Institute of Technology, GMR Institute of Technology, GMR Institute of Technology,
Rajam, Rajam, Vizianagaram Rajam, Vizianagaram

Abstract: positron emission tomography(PET) and so on. But

nowadays the process is of data gathering and parsing
Deep learning and convolutional neural networks are the the same to a developed deep learning model is a task
family of AI where the researchers had reached big success which could lead to generation of results which may or
using them in diagnosing. Over and above, a method is may not be as accurate as expected. This is just because
proposed to develop an AI model to assist radiologist to
improve diagnostic performance using convolutional
of the dataset chosen either categorical or continuous.
neural networks. This model is designed to work on Spect Now, the process of medical imaging (MI) came into
images, where the data has more features in 3D to consider the picture, which had changed the scope for easy
for gaining good accuracy compared with other identification of the thyroid disease or nodules and its
researchers works. stage or level at the foundation stage itself. Thee
systems can be developed on artificial intelligence
Keywords: Deep Learning, Convolutional Neural which helps the physician or radiologist to minimize the
Networks, Expert Advisory System, Medical Imaging, Spect costs and time in effective diagnoses (Chang Y et al.,
Images, Diagnosis and Prediction. (2016). Though several machine learning algorithms are
I.Introduction working on standard features of thyroid, the new emerge
Thyroid is one of the most common diseases and almost of Single-Photon Emission Computerized Tomography
42 million people in India suffer from it. One-third of (Spect) images can do better classification of malignant
the people unaware that they are affected with thyroid and benign tumour evolved because of thyroid.
and three among ten adults, are facing thyroid related 2. Literature Review
problems and issues. Most complicated is the role of
thyroid in pregnancy women’s, (especially goitre) which 2.1 International Status:
is leading to miscarriage because of the ungrown fetus.
Recent research, had given a confidence to computer Ma, L., Ma, C., Liu, Y., & Wang, X. (2019) presented
science professionals to provide expert advisory system that Spect imaging is practically important for the
(EAS) through machine learning ( Li LN et al., 2012) clinical diagnosis of thyroid diseases and explained the
which is the sub-components of Artificial Intelligence role of Spect images in the computer-aided diagnosis
(AI) to diagnose (Choi YJ et al., (2017) different kinds (CAD).
of diseases with high accuracy. The most common cause of goitre (Bryan Haugen et al.,
The medical professionals are made to use these 2013 ) is a lack of iodine in the diet. Iodine is a
systems due to some developed software’s during substance in food that the thyroid uses to make thyroid
general diagnosis process. Especially, the radiologists hormones. Growth of goitre and its troubling features
are playing the vital role in diagnosing the diseases by are most prompt elements in this study to rule and
using medical imaging like x-rays, computed understand goitre in 3D space.
tomography(CT), magnetic resonance imaging(MRI),
Lina Pedraza et al., 2015 researched saying that CAD nodules, whereas ultrasound elastography can claim
have been developed to assist radiologists in the and improve the sensitivity and diagnostic rate.
detection and diagnosis of abnormalities on the nodes
and nodules relevant to goitre.
The success of machine learning (Ker J., et al., 2018 )
algorithms at image recognition in recent years 2.2 National Status:
intersects the use of electronic medical imaging. This
diagnostic imaging is also named as medical imaging. Prasad et al., 2016-2021 as per the research, methods
General machine learning process like data pre- including Multi-Layer Perceptron (MLP), Radial Basis
processing is applicable on this categorized or Function (RBF) and Adaptive Conic Section Function
continuous data. Neural Network (ACSFNN) helped in diagnosis of TD;
the classification accuracies are 81to 88%. Similarly,
Theoretical analysis of generalization performance of Linear Discriminate Analysis (LDA), C4.5 with
feed forward networks requires training sample size to Default Learning Parameters (C4.5-1), C4.5 with
be linear in the number of adjustable parameters. These parameter c equal to 5 (C4.5-2), C4.5 with parameter c
requirements seemingly contradict empirical success of equal to 95 (C4.5-3) and DIMLP with two hidden
neural network applications where good generalization layers and default learning parameters (DIMLP) to
performance with finite training samples is often perform classification, and the accuracies reached 89%
achieved using large-size networks for thyroid cancer to 94% respectively.
and thyroid problems said Ahn HS et al., (2014).
Raghavendra U et al., 2018 focussed on ultrasound
Chen Q et al., 2010 proposed an image semantic imaging as one of common visualizing tools used by
classification approach based on Kernel PCA Support radiologists to identify the location of thyroid nodules.
Vector Machines (KPCA SVM). The KPCA, which is The evaluation of this CAD system on both private and
investigated from the complexity of optimization public datasets confirmed its effectiveness for the
problem and the generalization performance, is the radiological findings.
explicit extension of the optimal separating hyper planes
classifier. The experimental results on corel image However prompt, and valid accuracy is obtained with
dataset show that the proposed method is effective and the application of Artificial Immune Recognition
feasible. System (AIRS). Similarly, diagnosed thyroid diseases
expert system that called ESTDD (Expert System for
Mohamed cherif, Nait-Hamoud and Abdelouaheb Thyroid Disease Diagnosis) (Nikitha Singh et al., 2012),
Moussaoui (2010) compared the ability of the two novel whose accuracy was 95.33%. Finally, swarm
methods to discard outliers and enhance the optimization optimized support vector machines with
performance of the classification with PCA and FOAO fisher score (FS-PSO-SVM) CAD system for thyroid
and to highlight the efficiency of the combined method disease, and the average accuracy of 97.49% was
PCA-UC-FOAO in the classification of long-term ECG achieved.
records.
Major Findings in the Survey /Review
Xia S., et al., 2019 study proved that prospectively
included 180 thyroid nodules that underwent ultrasound  Thyroid classification by pathology is
interpretation. The radiologist assessed all nodules and common but it is a complex process.
the ultrasonographic features of different subtypes were Compared with international research, the
analysed, and the diagnostic performances of the CADs progress of thyroid diagnosis and its scope for
and radiologist were compared and found goitre nodules decision making is less in national research
are best for prediction. activities.
 All the previous research are dependent and
Juan J. Rodrı´guez et al., 2007 proposed a method for
executed on traditional data which is both
generating classifier ensembles based on feature
categorical and continuous.
extraction. It consists of splitting the feature set into K
subsets, running principal component analysis (PCA)  The scope for medical imaging measures is
always low in scope for applying the same for
separately on each subset and then reassembling a new
extracted feature set while keeping all the components. machine learning or deep learning algorithms
on image sets of TD.
Zhu LC et al., 2013 mainly aimed to construct a model  Imaging Technology is nowadays going fast
for using in differentiating benign and malignant in this developing world. The regularly used
nodules with the artificial neural network and to images are CT and Ultra Images. However,
increase the objective diagnostic accuracy. the focus is to be made on Spect images for
gaining accuracy.
Xia et al., 2017 claims the conventional ultrasound can  Though the data is consistent, sometimes the
be used as a routine examination technique in models refuse to load images which are above
differential diagnosis of benign and malignant thyroid
 forty thousand in size and whose resolutions the flow of blood, neurons active states and pulse could
are not in required format. be obtained in colour format.
 Comparison are done on various algorithms
which are implemented, which doesn’t In this proposed project the Spect images hold the
support same TDD with similar attributes on functional states of thyroid nodules through features
continuous datasets, image and video frame which are represented as important. The features
datasets. mainly focus on four categories a) Effect on fetus b)
Hypothyroidism c) Antibodies cross through Placenta
 Applying more algorithms increases the
efficiency of the cluster but it is not valid for a and d) Grave’s Disease. Besides this e) miscarriage at
earlier stages and f) fertility issues are considered to be
dataset which holds large knowledge base or
benign.
pixel representations.
 Testing sensitivity of algorithm vary from one The model is developed as shown in Figure 1, the
algorithm to another and model to model model is deployed in two phases which works on
which shows that all the research is required training and diagnoses. In the training phase, the
for clinical approval by the radiologists or the available ImageNet datasets are applied on various
concern doctors. models of machine learning and deep learning and will
check the accuracy parameters by passing the same
2.3 Importance of the current project in the context
through convolutional model over the network.
of current status
Similarly, in this training phase besides Data
The project could address all the pregnant ladies who
Augumentation(DA), the Transfer Learning,
are suffering with goitre. This project is exclusively
Optimization through Convolutional neural Networks
meant for pregnant women, because the growth of
also takes iterations to sustain the accuracy of the
thyroid gland and its hormone levels compared with
model. But while taking the thyroid Spect data the
the normal being are too high. Moreover, the scope of
images and the dataset is sent for DA. In this DA, the
getting the Spect images is possible for every
data is coupled based on the need by slight adjustment
preliminary check in their nine months pregnancy
of data by sample copying as the process. The obtained
period. The automated system is the software
sample data is processed and then the same is given as
application developed, which suggest the radiologist
the input to the improved training model. The scope of
about the dearness and level of thyroid during
how, the model is going to work is illustrated using
pregnancy, also provides the nearby referral source and
Figure 2.
produces the curing or prevention strategies to the
patient. The convolutional networks and AI are bought The obtained result is the resultant from the trained
into the study related to deeper prediction for diagnose model using any appropriate classification algorithms
at a narrow scope for examining the nodule states in of deep learning. Population based algorithm could
three-dimensional scope and understand the urgency react as the best algorithm for transfer learning
based on the features provided. The software would be algorithmic rate and speed to acquire good levels.
designed with combination of medical imaging and
convolutional neural networks via deep learning to The DA is processed as X aX1 + ( 1-a)X1 , where X
understand the effect of thyroid. is the resultant or the dependent variable of the sample
data which is iterated from the network and “a”
resembles the covariance of the dataset and “X1” is the
3. Objectives: target Spect data elements which are to be processed.
 To design and develop an AI model to assist
radiologist to improve diagnostic performance
using convolutional neural network.
 To classify the malignant and benign thyroid
nodules through Spect images.
 To feature scale and identify the pathway
involvement and validate the role of goitre
which in turn helps in tracing the scope for
miscarriage.
 To study the performance levels of the
proposed and existing models using generic
parameters.
3.1 Methodology:
AI had reported to meet the human expectations in
medical imaging. Spect Images uses gamma-ray
camera to obtain the dataset as needed. The beauty of
this Spect Images would arrange the data in 3D format.
This study is mainly focussed on Spect Images where
 Fig 1: Proposed Model

Later, the process is shifted on diagnosis where the

trained Spect images are applied as the input to the
Applied trained thyroid diagnosis model(ATTD) which
produces the classification of benign and malignant
tumour. The options a, b, c and d flows under
malignant tumours raised due to thyroid whereas the
options e, f are benign tumours. The diagnoses results
much focus on the most crucial problems identified in
the pregnancy.
Fig 3: RL and LL Feature Representation

Model Implementation

Fig2: Differentiation between Dense and Improve

Dense Block

The model processes the data based on the

augmentation with any of the techniques mentioned
like adding noise, cropping, clipping, rotation, scaling
and translation and so on. Transfer learning plays its
role in identifying the dense layers along with its
values before and after the training. Transfer learning
based on the features available, traces the most
prominent elements or features and combines them to Fig 4: Training and Fine Tuning the
mould as a single unit and reload them for obtaining Parameters through Improved Dense Layer
the output.Training Parameters are used to control the The balancing is to done through feature normalization
features for giving 100% accuracy for the resultant ( if needed) scaling to re-arrange the range of RL and
delivered as mentioned in Fig 4. LL. Rectified Linear Units (ReLU) is the formulation
is max(0,z)max(0,z). It is non-linear and provides the
same benefits as Sigmoid but with better performance.
Transfer learning always improves the efficiency by The focus is made on ReLU because it screens the 3D
maintaining the data accuracy at levels of every view of dataset pixel image. It completely remove
consistent dense layer. The dense layer image extra gradients and noise and normalize the data. Later,
processing rate may happen at 28 x 28 pixel or it may convolutional networks and convolutional activation
hold 56 x56 pixel frame. The input features of this function is raised to understand how accurately, the
project resultant as an input feature. Let, understand; model is classifying the tumours based on the goitre
how actually a Spect images looks like and what could strength. Probably the convolutional network based on
be the features in the image. Spect image is a collection the activation function insertion values, it transfers the
of siemen elements of the goitre having the right and data from one pixel to another as needed.
left locks as displayed in the image Figure 3.
 medical professional and further there is the scope for
the same to implement in other advanced technical
aspects, based on the datasets.
Where, is the function that is convoluted.
t represents the real number of functions f and g and
g(t) is the convolution function and (t~) is the first
derivative of the convolution function. Next, the depth References
concatenation would combine or concatenate the
elements of input feaatures with same height and  Ahn HS, Kim HJ, Welch HG 2014 Korea’s
width. Concatenation process mainly relies of the thyroid-cancer ‘‘epidemic’, screening and
trainable parameters and its outcomes after training. overdiagnosis. N Engl J Med, 371:1765–1767.
However, all the training parameters such as (kl, 0,  Anton Buhagiar. (2002), “Exploration and
k2,1, …., kl,l-1) understand the role of sample copying reduction of data using principal component
and accordingly samples the data for further analysis”, A journal of University of Malta
processing. Concatenation classes and layers are Medical School, Volume 14 , Issue 1 ,pp: 27-
deployed in pytorch or python using Keras to obtain 35.
list of Inputs with same features to produce the outputs  Asteria C, Giovanardi A, Pizzocaro A,
of its kind keeping all features as generic parameters. Cozzaglio L, Morabito A, Somalvico F 2008
The trainable parameters would show the impact of US-elastography in the differential diagnosis of
existing data on the improved model which is obtained benign and malignant thyroid nodules. Thyroid
through sampling. The limitations took over sigmoid 18:523-531.
and Tanh functions had added a scope for ReLU to  Bryan Haugen, MD, James Hennessey, MD,
work with social netwoking models to predict the Leonard Wartofsky, MD, Goiter, The Journal
outputs. However, the Rectified Linear Activation of Clinical Endocrinology & Metabolism,
function changes it value layer by layer based on the Volume 98, Issue 1, 1 January 2013, Pages
dense layer properties. The extensions to this ReLu 27A–28A,
function can also add limitless accuracy standards.  Chang Y, Paul AK, Kim N, Baek JH, Choi YJ,
The training parameters more focus on the datasets and Ha EJ, Lee KD, Lee HS, Shin D, Kim N 2016
how those values are trained, fo example the the Computer-aided diagnosis for classifying
enhanced data after sample copying is processed in benign versus malignant thyroid nodules based
training phase, the local data is also processed in on ultrasound images: a comparison with
training phase whereas the performance data is  Chen, Q., Chen, X., & Wu, Y. (2010).
processed in unlabbelled format in the design phase. At Optimization Algorithm with Kernel PCA to
the outset, this Figure 4 completely focuses on all the Support Vector Machines for Time Series
features and finally end up with the most promising Prediction. JCP, 5, 380-387.
features after successful training and comparison.  David R. Hardoon, Sandor R. Szedmak, and
The improved dense block certainly increases the spect John R. Shawe-taylor. (2004). “Canonical
images digital followup to classify thyroid. The most Correlation Analysis: An Overview with
techincal components which evolves as the output is i) Application to Learning Methods”. Neural
Classification and Performance evaluation ii) Average Computing. Vol. 16, Issue. 12, pp: 2639–2664.
Precission and Efficient Iterations iii) Confusion  Gu, G., Liu, H., Shen, J., & Shi, L. (2008).
Matrix with perfect Accuracy and if needed iv) a Image Semantic Classification algorithm
Comparitive study can also be proposed to check the Research On Kernel PCA support vector
scope of its uniqueness compared with other models. machine. 2008 IEEE International Symposium
on Knowledge Acquisition and Modeling
4. Conclusion and Future Scope Workshop, 422-424.
 Goitre Open Access Datasets (2015) :
The research outcome of this project will be utilized in
http://cimalab.intec.co/applications/thyroid/
medical field especially by the radiologist and doctors
 H. Huang, H. He, X. Fan, and J. Zhang (2010),
to predict the level of thyroid at every stage of
“Super-resolution of human face image using
pregnancy. This proposal can act as a human
canonical correlation analysis,” Pattern
surveillance to get rid from TD using a computational
Recognition., Vol. 43, No. 7, pp. 2532–2543.
approach via deep learning. This project leads itself in
identification, detection, and suggestions at preliminary  Hastie T, Tibshirani R, Friedman J 2009 The
stages of pregnancy to avoid miscarriage. The software Elements of Statistical Learning. Second
application developed can be used in the TD-CAD edition. Springer, New York, NY.
centres and could act as a cache reference for the  J. Zhou, Y. Lin and Y. Chen, "Ensemble
doctors or medicinal experts. The infrastructure Classifiers based on Kernel ICA for Cancer
procured during the period of research findings can be Data Classification," 2009 2nd International
further used for updating and maintaining the database Conference on Biomedical Engineering and
with newly generated reports and datasets given by the Informatics, Tianjin, 2009, pp. 1-5.
 conference on Emerging intelligent computing nodules: an extreme learning machine
technology and applications (ICIC’09). approach. Comput Methods Programs Biomed
Springer-Verlag, Berlin, Heidelberg, 955–964. 147:37–49.
 Ker J., Wang L., Rao J., Lim T. Deep learning  Xia, S., Yao, J., Zhou, W. et al. A computer-
applications in medical image analysis. IEEE aided diagnosing system in the evaluation of
Access. 2018;6(1):9375–9389. thyroid nodulesexperience in a specialized
 Ludmila I. Kuncheva1 and Juan J. Rodr´ıguez2 thyroid center. World J Surg Onc 17, 210
(2007), “An Experimental Study on Rotation (2019).
Forest Ensembles”, Springer, pp: 459-468  Zhu LC, Ye YL, Luo WH, Su M, Wei HP,
 Lina Pedraza, Carlos Vargas, Fabián Narváez, Zhang XB, Wei J, Zou CL 2013 A model to
Oscar Durán, Emma Muñoz, Eduardo Romero, discriminate malignant from benign thyroid
"An open access thyroid ultrasound image nodules using artificial neural network.
database," Proc. SPIE 9287, 10th International
Symposium on Medical Information
Processing and Analysis, 92870W (28 January
2015
 Ma, L., Ma, C., Liu, Y., & Wang, X. (2019).
Thyroid Diagnosis from SPECT Images Using
Convolutional Neural Network with
Optimization. Computational intelligence and
neuroscience, 2019, 6212759.
 Misiakos EP, Margari N, Meristoudis C,
Machairas N, Schizas D, Petropoulos K,
Spathis A, Karakitsos P, Machairas A 2016
Cytopathologic diagnosis of fine needle
aspiration biopsies of thyroid nodules. World J
Clin Cases 4:38–48.
 Nait-Hamoud, M.C., & Moussaoui, A. (2010).
Two novel methods for multiclass ECG
arrhythmias classification based on PCA, fuzzy
support vector machine and unbalanced
clustering. 2010 International Conference on
Machine and Web Intelligence, 140-145.
 Nikitha Singh , Alka Jindal(2012), “A
segmentation method and classification of
diagnoses for thyroid nodules”, IOSRJCE,
Vol.1, Issue 6, August ,pp:22-27.
 Raue F, Frank-Raue K 2016 Thyroid cancer:
risk-stratified management and individualized
therapy. Clin Cancer Res22:5012–5021.
 Raghavendra U., Gudigar A., Maithri M., et al.
Optimized multi-level elongated quinary
patterns for the assessment of thyroid nodules
in ultrasound images. Computers in Biology
and Medicine. 2018;95(1):55–62.
 Thyroid disease records
https://archive.ics.uci.edu/ml/datasets/Thyroid+
Disease
 Vanden Broeck J, Cunningham SA, Eeckels R,
Herbst K 2015 Data cleaning: detecting,
diagnosing, and editing data abnormalities.
PLoS Med 2:e267.
 Vladimir Cherkassky(2017), “Model
complexity control and statistical learning
theory”, Natural Computing, Springer , Vol 1,
pp: 109-133.
 Xia J, Chen H, Li Q, Zhou M, Chen L, Cai Z,
Fang Y, ZhouH 2017 Ultrasound-based
differentiation of malignant and benign thyroid