VITAMIN DEFICIENCY DETECTION USING

2024 International Conference on Wireless Communications Signal Processing and Networking (WiSPNET) | 979-8-3503-5084-5/24/$31.00 ©2024 IEEE | DOI: 10.1109/WISPNET61464.2024.10533044

NEURAL NETWORKS

T.Krishna Chaithanya[1], Jiniya Gupta[2] ,M.S.Roobini[3]

[email protected] [1], [email protected] [2] ,[email protected] [3],
Department of computer science and engineering [12,3], Sathyabama Institute of Science and Technology [1,2,3]
Chennai-600119, India

Abstract—Vitamin deficiency is a pervasive health issue affecting Vitamin deficiencies highlight the many health problems we
millions globally, often leading to severe health complications if face every day. Vitamin deficiencies occur from insufficient
undiagnosed. Various vitamin deficiencies can be identified by
identifiable symptoms manifesting in numerous areas across the intake of vitamins. Eating a balanced diet is good for general
human body. Blood tests are most commonly used to detect health. Dietary imbalances can lead to excess or insufficient
deficiencies. However, blood tests are expensive and have limited intake of certain nutrients. Deficiencies can occur when you
availability. The project aims to develop an automated system to
don't get enough of a particular nutrient. Vitamins are
detect vitamin deficiencies using Mobilenet, convolutional neural
networks (CNN), NasNet-mobile, inception modules, and artificial important micronutrients that the body requires for various
neural networks (ANN). The program allows users to detect physiological processes. An insufficiency of vitamins results
potential vitamin deficiencies without needing blood samples by in various health problems, from fatigue and poor immunity
examining images of the eyes, lips, tongue, and nails. Here we
have considered the dataset of abnormalities caused by vitamin
to severe conditions like scurvy, rickets, and Anemia. A
deficiencies like glaucoma eyes, bluish nails, and other nutritionally deficient diet can lead to a variety of symptoms.
abnormalities of lips and tongue. We have trained the model on The body uses these symptoms as a warning system for
these image datasets, by leveraging image recognition and data possible nutritional or mineral deficiencies. Knowing these
analytics, the proposed system offers a quick, efficient, and non-
invasive method of detection thereby streamlining the diagnostic will facilitate better management of your diet. Deficiency of
process and facilitating early intervention. The aim is to provide certain nutrients in the body determines the symptoms of
an economical, accessible, and efficient clinical screening method. deficiency in biological processes. Early detection of vitamin
Moreover, the platform furnishes tailored dietary suggestions to
deficiencies makes the user’s work easier and saves time and
tackle pinpointed inadequacies, thus alleviating health hazards
linked with insufficient nourishment. By furnishing an easily money. Blood tests are standard practice in rich countries, but
reachable resource for prompt identification and action, our different socio-economic and public perspectives make
system aids in tackling a prevalent worldwide health concern. The testing difficult in poor countries. Traditional methods of
program functions as a beneficial aid for individuals to combat a
widespread global challenge, impacting millions due to inadequate detecting vitamin deficiencies often involve blood tests and
comprehension of nutrition. Serving as a valuable resource, it symptom-based clinical diagnoses, which can be invasive,
addresses an international issue, assisting those facing nutritional costly, and time-consuming. Micronutrient deficiencies are a
uncertainties worldwide.
serious global health problem. The estimated survey says that
around 20 crore individuals globally experience deficiencies
in nutrients like zinc, iron, iodine, and vitamin A. “More than
Keywords— Vitamin deficiency, CNN, NASNet-mobile, 1.2 billion individuals are Zinc deficient and half a million of
Inception, non-invasive. them die each year. Similarly speaking, over 100,000 people
die due to Anemia caused by iron deficiency” [1].
I. INTRODUCTION “Throughout the world, poor people do not consume
sufficient amounts of nutrient-rich foods such as meat, eggs,
Micronutrients that the body does not produce in sufficient fish, milk, legumes, fruits and vegetables. The problem is
amounts are vitamins. This explains why the body needs to made worse by inadequate health care and sanitation, disease,
receive nutrients from external sources to function properly. and a lack of education in infants and childcare” [17]. Most

979-8-3503-5084-5/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 individuals facing these deficiencies reside in developing providers often suggest blood tests to assess the levels of
nations and frequently lack multiple micronutrients. Lack of vitamins D, B12, and A when patients exhibit symptoms such
access to foods high in vitamins leads to deficiencies. This as weakness, fatigue, or skin issues. However, in developing
restriction is frequently caused by the high price or scarcity nations in Africa and Southeast Asia, blood tests are more
of particular items in the community. Micronutrient deficits difficult to get and more expensive, which makes it more
raise the danger of contracting illnesses and the mortality rate difficult to identify vitamin A deficiency in time to stop
of malaria, measles, pneumonia, and diarrhea. These illnesses further problems [2]. This process entails a comprehensive
are among the top 10 most common reasons of disease clinical evaluation, where the doctor conducts a physical
globally. “The risks are higher in pregnant women, children examination and considers the patient's medical history.
(from conception to young childhood), adolescents, the Subsequently, a blood sample is collected and sent to a
elderly, and all of the over 800 million people globally who laboratory for thorough analysis. Based on the results of both
are undernourished. At risk are approximately 125 million the lab tests and the clinical evaluation, clinicians can
preschool children with vitamin A deficiency, as well as sub- formulate a diagnosis. Nevertheless, this traditional approach
populations at risk of deficiencies of folate, thiamin, vitamin comes with notable drawbacks. The invasive nature of blood
B12, niacin, riboflavin, and other B vitamins and vitamin D” tests, involving the insertion of a needle into a vein, poses a
[12]. A country cannot reach its full potential if its whole slight risk of bruising and may induce discomfort or anxiety
population suffers from malnutrition, including shortages in in certain individuals. Additionally, the process is time-
vital vitamins and minerals. This leads to higher healthcare consuming, both during the actual test and the subsequent
costs, stalled educational programs, a workforce that is less wait for results, potentially causing delays in the diagnosis
skilled and productive, and a decline in total economic and treatment. The high costs associated with blood tests
activity. As a result, human capital is drastically reduced [17]. further restrict accessibility, particularly for individuals
When vitamin deficiencies occur due to insufficient intake, residing in impoverished or rural areas where access to
these are classified as primary deficiencies. When an specialists and specialized equipment is often limited.
underlying disease, such as malabsorption, causes deficiency Exploring alternative diagnostic methods that are less
symptoms, they are called secondary deficiency symptoms. invasive and more accessible is crucial in addressing the need
The underlying condition may be metabolic, such as a genetic for detecting vitamin deficiencies.
defect in the conversion of tryptophan to niacin. It may also
be due to lifestyle choices that increase your vitamin needs,
III. LITERATURE SURVEY
such as drinking alcohol or smoking. According to
government standards for vitamin deficiencies, healthy
people should consume certain amounts, and certain amounts
are recommended for women, men, newborns, the elderly, In this paper [1] researchers extracted certain characteristics
and pregnant or lactating women. The term “vitamin and qualities from the photos, and the program used machine
deficiency” is often used today. This expression is directly learning to make decisions for determining the kind of
related to the state of a person's health due to daily activities shortfall with a fuzzy logic. This research introduces a free
and habits. This study focuses on the use of all techniques and artificial intelligence program that uses certain bodily organs
methods used in the diagnosis of vitamin deficiencies. To to identify vitamin deficits in individuals. Researchers have
address common vitamin deficiencies, many countries have developed an Android mobile application equipped with
implemented mandatory vitamin fortification programs. An artificial intelligence that can diagnose a range of vitamin
unstable lifestyle and the habit of eating bad foods can lead deficiencies by analyzing user-provided photos of nails, lips,
to vitamin deficiencies. Over time, a person's expertise and tongue, and eyes [1]. In this study, researchers explored “the
skills can deteriorate. The utilization of deep learning (DL) potential of utilizing Machine Learning and Explainable
and machine learning (ML) methods to improve clinical Artificial Intelligence methods to provide a more accessible
decision-making is gaining interest in the health sciences and system for Vitamin A deficiency detection using Electrical
medical field. This software is a tool that helps address the Health Records of symptoms and a variety of machine-
global problem of low nutritional awareness and helps learning techniques are applied to a sparse dataset of ocular
medical professionals make more accurate diagnoses. symptoms and diagnoses” [2].In this paper, researchers
designed a technology to find hidden trends and eating habits
of clients using a variety of data sources. This method will
II. EXISTING SYSTEM help monitor and enhance a person's health and identify the
kinds of food they should stay away from to lower their risk
Presently utilized diagnostic frameworks primarily detect of illness. The research helps in both predicting a healthy diet
potential vitamin deficiencies through biochemical testing, for each individual and creating a diet plan tailored to the
particularly blood tests, and clinical evaluations. Healthcare patient's requirements. [3].

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 In this paper [4], researchers used the kind of deficit thereby saving time and costs while ensuring user-friendly
specified using the fuzzy logic approach. Several annotated accessibility. Following each diagnostic method, the
photos of segmentation symptoms were used to train a application provides recommended medications and suggests
TensorFlow classifier after pathological research had appropriate dietary measures accordingly [11]. In this study
specified the visual symptoms. The program is taught to [12], The researchers comprehensively elucidated the latest
discriminate between user photographs of the nails, tongue, information regarding vitamin deficiencies and underscored
eyes, and lips of normal persons of those who have a public health strategies to effectively tackle these issues.
deficiency in vitamins.In this paper researchers performed by Furthermore, they emphasized the significance of addressing
contrasting the DCNNs' performance with a support vector micronutrient deficiencies, emphasizing the need to identify
machine's color feature and oriented gradient histogram, and individuals at risk and implement proactive measures for
the DCNNs were shown to be effective. This work shows that prevention and management. The researchers highlighted the
DCNNs are a useful method for identifying nutritional imperative nature of a holistic approach in addressing these
deficits in rice. The author explored the accuracy of various nutritional gaps to safeguard public health. In this study,
Deep Convolutional Neural Networks in detecting nutritional researchers explained how meticulously carried out clinical
deficiencies inside rice within the scope of this research [5]. investigations and Field-based randomized trials, altering
global health policy, unveiled the complete extent and
In this research, scientists have introduced an expert system repercussions of insufficiency among impoverished
application designed to assist the community in identifying populations in developing countries [13]. In this study, the
early-stage vitamin deficiencies. This innovation not only researchers explained current knowledge about the kinetics
saves time and expenses but also simplifies the process for of different forms of vitamin K, their detection, and their
both users and researchers. The researchers elaborate that the toxicity are discussed [14].
field of artificial intelligence, particularly through expert
systems, enables the early detection of vitamin and mineral
deficiencies. Likewise, by discerning precise symptoms, they The researchers of this study [15] explained the effectiveness
can accurately identify the specific vitamins and minerals that of DCNNs was affirmed through comparison with color
are lacking in the body [6]. The objective of the researchers features using a support vector machine and a histogram of
is to create a screening system for Vitamin A deficiency in oriented gradients employing a support vector machine. Their
children that eliminates the need for retinol serum blood tests. investigation illustrates that DCNNs offer a potent
Instead, the system utilizes available health records. The methodology for the diagnosis of nutrient deficiencies in rice.
Support Vector Classifier model demonstrated the highest In this study [16], The findings presented by the researcher
accuracy scores, with an accuracy of 75.7%, sensitivity of suggest that when evaluating patients exhibiting severe dry
83.7%, and specificity of 74.9%. [7]. In this investigation, the eye and corneal ulceration, it is crucial to include vitamin A
scientists employed RCNN, an exceptionally sophisticated deficiency resulting from eating disorders in the list of
machine learning algorithm known for its superior potential diagnoses. The micronutrients zinc, iron, iodine,
performance across diverse pattern recognition tasks, making folate, and vitamin A are necessary for human health and
it a fitting selection for this particular application. The growth. This paper [17] delves into the repercussions of
primary aim of this research endeavor is to make a significant deficiencies in crucial micronutrients on human health,
contribution to the domain of dermatology by elevating the provides recommendations for implementing micronutrient
precision in diagnosing vitamin deficiencies and amplifying programs, and demonstrates the substantial influence that
the effectiveness of therapeutic interventions through the investments in this realm can have on enhancing human
utilization of state-of-the-art imaging technology [8]. capital. Furthermore, it underscores the imperative need for
In this research, the dataset is enriched through augmentation sustained efforts and global collaboration in addressing
employing linear transformations. Subsequently, the micronutrient deficiencies to promote long-term public
augmented images undergo preprocessing in multiple color health and well-being.
spaces before being introduced to the neural network [9].
The objective of this paper is a deliberate endeavor to
establish an autonomous and reliable cost-effective solution
IV. PROPOSED SYSTEM
for detecting nutritional deficits. Datasets representing
deficient and healthier categories are formulated employing
image processing techniques such as an RGB color feature The suggested approach in this research seeks to identify
extractor, edge detection, real-time texture recognition, and vitamin deficiencies through the use of deep learning models.
similar methodologies [10]. In this paper, the researchers For correct analysis, it makes use of NASNetMobile,
developed an expert system application designed to assist the Convolutional Neural Networks (CNN), Inception, Artificial
community in early diagnosis of vitamin deficiencies, Neural Networks (ANN), and mobile net. Numerous benefits

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 come with the suggested method, including its accessibility, reach an optimistic level of accuracy. We will save the
efficiency, high scalability, and non-invasive nature. accuracy in the backend and move on to another model as
soon as we have it.
A. Objective

The objective is to create an extensive dataset that includes a

V.METHODOLOGY
range of symptoms and visual cues related to vitamin A. Dataset
deficiency. To use CNN, ANN, Inception, and
NASNetMobile to construct and train a reliable neural In this investigation, the dataset encompasses a diverse set of
network model. to have a minimum 90% accuracy rate in 2,000 images designated for training and 400 images
identifying vitamin deficits. to incorporate the model into an earmarked for testing purposes. These images span a wide
approachable application for testing and validation in the real spectrum of visual characteristics, encompassing distinctive
world. features such as bluish nails, eyes affected by glaucoma, and
intricate depictions of the lip and tongue. The dataset has
B. System architecture been meticulously curated to include a variety of attributes
associated with potential vitamin deficiencies in humans. The
incorporation of diverse visual elements is intended to
augment the model's capacity to discern and categorize
IMAGE IMAGE MODEL
PREPROCESSING TRAINING symptoms indicative of vitamin deficiencies. Importantly, all
DATASET
images employed in this dataset have been responsibly
gathered from publicly accessible repositories, with a
primary focus on Google Photos. This sourcing strategy
ensures a comprehensive representation of authentic real-
Bluish nails,
glaucoma eyes, lips,
world cases, contributing significantly to the diversity and
tongue authenticity of the dataset. This approach is geared towards
facilitating the development of robust classification models
CLASSIFICATION MODEL
EVALUATION adept at recognizing and addressing various presentations of
vitamin deficiencies within diverse populations.

DEPLOYMENT

Fig 4.1: System Architecture

Fig 4.1 explains the system architecture and the process.

Gathering data is the initial stage in every deep learning
technique. First, information is gathered via photos. These
might be pictures of symptoms related to the eyes, tongue, Fig 5.1 Sample dataset
lips, or nails found online or in Google Photos. Following
dataset collection, pre-processing is carried out. Fig. 5.1 shows the sample dataset consisting of 4 categories
Preprocessing is done on data to prepare it for analysis. that are glaucoma eyes, abnormal conditions of the tongue,
and lips and bluish nails.
For example, data may be encoded and images may be
shrunk. Pre-processing includes feature extraction, which B. Data Preprocessing
involves removing pertinent traits such as blue nails and
glaucoma-affected eyes. Subsequently, the dataset is The image dataset undergoes initial preprocessing, including
partitioned into training and testing subsets. The model has resizing for standardized dimensions and pixel rescaling.
been constructed, and it will be trained using training 1. Resizing: Images are resized to 224x224 for ANN, CNN,
datasets. In addition, we will supply testing datasets for the NasNet, and MobileNet models and 256x256 for the
model to be tested, or the user can submit a test image Inceptionv3 model.2. Image Rescaling: Each image
directly. Following training cycles and epochs, the model will undergoes pixel rescaling to a range between 0 and 1 by

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 dividing pixel values by 255, simplifying computational InceptionV3 model demonstrates exceptional proficiency in
processes. mastering intricate hierarchical representations within image
data. This capability, coupled with training for 10 epochs,
C. Models signifies its robustness in handling complex visual data.

1. ANN Model 4.NASNetMobile

During the training stage, a dataset containing 2,000 At the core of this architecture lies the
images spread across four classes was employed to train a NASNetMobile model, which is trained in advance on the
sequential neural network model. The model structure, ImageNet dataset. Tailored for mobile applications, the
described as a sequential arrangement of dense layers, model strikes a delicate balance between computational
encompasses a flattened layer for reshaping input data, efficiency and high-performance image classification. To
succeeded by three densely connected layers featuring 128, capitalize on the acquired knowledge from pre-training, the
64, and 32 units, respectively. The last dense layer produces layers of the base model are fixed, ensuring the retention of
predictions for the four classes. The overall parameter count learned features throughout subsequent training stages.
for the model stands at 19,278,180 (73.54 MB), with all The custom classification head complements the base model
parameters open for training. This particular model by integrating global average pooling for dimensionality
underwent 30 epochs, iterating through the dataset 30 times, reduction, a dense layer featuring 1024 units with ReLU
aiming to optimize its weights and biases and improve activation for effective feature extraction, and an optional
performance on the training dataset. dropout layer for regularization. The final dense layer,
utilizing softmax activation, produces predictions for the
2. CNN Model designated four output classes. Notably, this composite
NASNetMobile-based architecture is fine-tuned for the
The CNN model, designed with a sequential specific task, and the entire model undergoes training for 15
architecture, excels in capturing intricate patterns within epochs to refine its performance and elevate its effectiveness
image data. The first convolutional layer (conv2d) is in the designated classification task.
composed of 512 filters, succeeded by a subsequent layer of
max-pooling (max_pooling2d) for feature extraction and 5.MobileNet
spatial dimension reduction. The subsequent convolutional
layer (conv2d_1), featuring 256 filters, and its corresponding The MobileNet architecture, utilized as the
max-pooling layer (max_pooling2d_1) further refine feature foundational model in this setup, has undergone pre-training
abstraction. The flattened layer prepares the output for on the ImageNet dataset, specifically designed for mobile
densely connected layers. Two dense layers (dense and applications. With an input shape of (224, 224, 3), the base
dense_1) with 64 and 4 units, respectively, contribute to the model's layers, preserving knowledge acquired during pre-
final classification. With a total of 9,124,420 trainable training, function as a feature extractor. The subsequent
parameters (34.81 MB), The model underwent training for a design encompasses global average pooling for reducing
total of 10 epochs. dimensionality, succeeded by three densely connected layers.
The initial two dense layers, each incorporating 1024 units
3.InceptionV3 activated by ReLU, augment the model's ability to enhance
the model's capacity to comprehend complex functions and
The sequential InceptionV3 model adopts a multi- classify intricate patterns. The third dense layer, featuring 512
tiered architecture for effective feature extraction and units, further enhances the model's capabilities in
hierarchical representation learning. The initial convolutional representing features. The ultimate dense layer, employing
layers, conv2d_4 and conv2d_5, each featuring 32 filters, are softmax activation, produces predictions for the designated
followed by a max-pooling operation (max_pooling2d_2) to four output classes. It is noteworthy that this MobileNet-
condense spatial dimensions. Subsequent convolutional based architecture is subjected to training for 10 epochs,
layers, conv2d_6 and conv2d_7, each equipped with 64 undergoing iterative optimization over the dataset to refine its
filters, further extend feature abstraction, culminating in an parameters and enhance its performance in the assigned
additional max-pooling step (max_pooling2d_3). The classification task.
flattening layer prepares the output for densely connected
layers, where dense_2, consisting of 256 units, facilitates
intricate feature mapping. The final dense layer, dense_3, D. Optimizer
with 2 units, enables classification. With a total parameter The Adam optimizer, which combines the advantages of
count of 67,193,122 (256.32 MB), entirely trainable, the momentum and root mean square propagation, represents a

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 widely used optimization algorithm in the realm of deep Fig. 6.2 shows The model undergoes training for 10 epochs,
learning. It allows for the dynamic adjustment of learning during which a concurrence between training accuracy and
rates for every parameter, which promotes robust validation accuracy is observed in the CNN model. The final
performance and effective convergence when neural train and validation accuracies are 95.7% and 96.2%.
networks are being trained. In the context of this study, Adam
optimizer has been consistently employed across all models, 3.INCEPTIONV3
ensuring a standardized and well-performing optimization
approach throughout the training processes.

VI. RESULTS AND DISCUSSIONS

In this investigation, an in-depth examination of the training
and validation performance of diverse models, including the
Convolutional Neural Network (CNN), InceptionV3,
MobileNet, Artificial Neural Network (ANN), and
NASNetMobile, has been conducted. The assessment is
carried out by analyzing accuracy and loss graphs, generated
Fig 6.3: Accuracy graph of INCEPTIONV3
from the training history of each model. Crucial metrics such
as training loss, validation loss, training accuracy, and
Fig. 6.3 shows the InceptionV3 model has poor agreement
validation accuracy are meticulously derived from the history
between training and validation accuracy. After the 3rd epoch,
object, offering a comprehensive understanding of the
the validation accuracy didn’t improve and the final training
learning dynamics exhibited by each model during the
and validation accuracies were 98.5% and 79.1%
training process.
A. Training and Validation Accuracy History 4. NASNETMOBILE

1. ANN Model

Fig 6.4: Accuracy graph of NASNETMOBILE

Fig 6.1 Accuracy graph of the ANN model

Fig 6.4 shows the NASNetMobile model is trained for 15

Fig. 6.1 shows the training accuracy increased till the 7th
epochs and during the 15th epoch the training and validation
epoch but slightly decreased after it. Then the accuracy
accuracies of the model are 95.8% and 96.5%.
steadily increased and reached 89.2%. The final validation
accuracy is 93%.
5. MOBILENET
2. CNN Model

Fig 6.2: Accuracy graph of CNN model Fig 6.5: Accuracy graph of MOBILENET MODEL

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 Fig. 6.5 shows the MobileNet model has shown better 3. INCEPTIONV3
agreement between training and validation accuracies among
all the models. The MobileNet model is trained for 10 epochs
and the final training and validation accuracies of the
MobileNet model are 96% and 97.2%.

B. Training and Validation Loss History:

1. ANN Model

Fig 6.8: Loss history of INCEPTIONV3

In Fig. 6.8, the final validation loss is 0.049 and the final
training loss is 1.115 which shows a clear case of overfitting.

4. NASNETMOBILE

Fig 6.6: Loss history of ANN model

Fig. 6.6 shows the ANN model has been trained for 30 epochs
and the final training and validation losses of the model are
0.2716 and 0.1853.

2. CNN Model

Fig 6.9: Loss history of NASNETMOBILE

In Fig. 6.9, the 1st epoch, the training and validation losses are
0.3195 and 0.111. During the 15th epoch, the losses of the
trained model are 01038. and 0.0662.

5. MOBILENET

Fig 6.7: Loss history OF CNN model

In Fig. 6.7 the first epoch, the training and validation losses
of the model are 1.4216 and 1.0684 after training for 10
epochs the losses of the trained model are 0.1053 and 0.0619. Fig 6.10: Loss history of MOBILENET

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 The above Fig. 6.10 shows the MobileNet model, the 1. MobileNet Confusion Matrix:
validation loss is 1.111, which is very high compared to the
training loss. After training the model for 10 epochs the
validation loss has come down to 0.074 and the final training
loss is 0.0940, this shows the model's ability to learn real
unlabelled data.

C. Performance Analysis:

Table 1 shows the performance analysis and among

the models assessed, the Convolutional Neural Network
(CNN) emerges as notably effective, attaining impressive
training accuracy of 95.7%, highlighting its superiority over
the Artificial Neural Network (ANN). Incorporating
convolutional layers in CNN enables the extraction of
hierarchical features from image data, contributing to its
enhanced performance compared to the ANN, which Fig 6.11 Confusion Matrix of MobileNet
demonstrates a slightly lower training accuracy of 89.2% and
a validation accuracy of 93%. In the Fig 6.11 Confusion matrix of MobileNet, the Class
bluish nail had 108 correct predictions, and the Class
However, MobileNet outperforms all others, glaucoma eyes had 100 correct predictions. Class lip had 89
achieving the highest overall accuracy at 97.2% in the correct predictions, with 11 instances misclassified as the
validation set, showcasing its adeptness in generalizing well tongue. The class tongue had 100 correct predictions, with 0
to unseen data. Conversely, InceptionV3, with a high training instances misclassified.
accuracy of 98.5%, displays a notable disparity with a
validation accuracy of 79.1%, indicating a potential case of 2.NasNetMobile Confusion Matrix
overfitting. These comparative insights underscore the
nuanced performance of each model, emphasizing the need
for a comprehensive evaluation across training and validation
metrics to assess their efficacy in real-world applications.

Table 1: PERFORMANCE ANALYSIS

Model Training Validation

Accuracy Accuracy

ANN 89.2% 93%

CNN 95.7% 96.2%

Fig 6.12 Confusion Matrix of NasNetMobile

InceptionV3 98.5% 79.1%

In the Fig. 6.12 Confusion matrix of NasnetMobile, the Class

NASNetMobile 95.8% 96.5% bluish nail had 108 correct predictions, and the Class
glaucoma eyes had 100 correct predictions. Class lip had 95
MobileNet 96% 97.2% correct predictions, with 5 instances misclassified as the
tongue. The class tongue had 98 correct predictions, with 2
instances misclassified. The above matrices provide a
Table 1, shows the performance analysis of various deep detailed overview of the model's performance across different
learning models using accuracy metrics. classes, helping to identify where the model excels and where
it struggles in terms of classification.
D. Confusion Matrix

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 E. Model Deployment VIII. CONCLUSION

The model has been trained on a dataset containing

images of glaucoma eyes, as well as abnormalities in the lips
and tongue, and bluish nails caused by respective vitamin
deficiencies. The neural network models have achieved good
accuracy in identifying these abnormalities, aiding in the real-
time detection of new data. This application is a novel method
that enables quick self-diagnosis without requiring a blood
sample. It serves as a means to heighten community
awareness regarding unaddressed nutritional requirements
and assist in obtaining an appropriate diet, not a substitute for
medical advice. This will help prevent untreated vitamin
deficiencies from leading to severe health concerns.
Nevertheless, because of limited access to photographs of
Fig 6.12 Deployment of the model people with vitamin deficiencies, the implementation was not
deployed on actual diseased persons. By adding data and
In Fig 6.12 the deployment phase of my Vitamin Deficiency directly involving experts, researchers, and doctors in the
Classification initiative, I employed the Django framework database, there is potential for a substantial enhancement in
for the backend, complemented by HTML, CSS, and diagnostic accuracy.
JavaScript for the front end, all orchestrated using Python.
The operational architecture integrates a selection
REFERENCES
mechanism among five distinct models, namely CNN, ANN,
NasNetmobile, MobileNet, and InceptionV3, facilitating [1] A. S. Eldeen, M. AitGacem, S. Alghlayini, W. Shehieb and M. Mir,
diverse predictive analyses. Following the model selection "Vitamin Deficiency Detection Using Image Processing and Neural
process, image standardization to dimensions of 224x224 is Network," 2020 Advances in Science and Engineering Technology
International Conferences (ASET), Dubai, United Arab Emirates,
undertaken before the application of the model.predict() 2020, pp. 1-5, doi: 10.1109/ASET48392.2020.9118303.
function. The resulting output not only discerns symptoms [2] D. A. Abuhani, J. Khan and H. Sulieman, "Detecting Vitamin A
such as bluish nails, lip irregularities, tongue anomalies, or Deficiency in Schoolchildren Using an Enhanced Explainable Machine
Learning Model," 2023 Advances in Science and Engineering
glaucoma eyes but also presents associated precautions and Technology International Conferences (ASET), Dubai, United Arab
identifies the specific type of deficiency corresponding to the Emirates, 2023, pp. 1-5, doi: 10.1109/ASET56582.2023.10180556.
observed symptomatology. [3] N. Naga Lakshmi, M. Jagadeesh Reddy, K. Hari Krishna and S.
Sindhuja Reddy, “Vitamin Deficiency and Food Recommendation
System Using Machine Learning,” 2022 IJRASET, doi:
https://doi.org/10.22214/ijraset.2022.43236
VII. FUTURE ENHANCEMENT [4] Dr. Arati Dandavate, Priyanka Gore, Namita Naikwadi, Shrushti Sable,
Muskan Tilwani, “Vitamin Deficiency Detection Using Image
Processing and Artificial Intelligence,” 2021 International Research
In the prospective scope of this research, there is Journal of Engineering and Technology
substantial potential for improvement. While the dataset was (IRJET),https://www.irjet.net/archives/V8/i4/IRJET-V8I4642.pdf.
meticulously sourced from Google Photos, upcoming studies [5] Zhe Xu, Xi Guo, Anfan Zhu, Xiaolin He, Xiaomin Zhao, Yi Han,
Roshan Subedi, "Using Deep Convolutional Neural Networks for
stand to gain by incorporating images from varied and Image-Based Diagnosis of Nutrient Deficiencies in
credible sources such as medical platforms and hospitals. Rice", Computational Intelligence and Neuroscience, vol. 2020,
This extension aims to construct a more comprehensive Article ID 7307252, 12 pages, 2020.
https://doi.org/10.1155/2020/7307252,
dataset, enhancing the model's proficiency in identifying [6] S. Yadav, A. Rathod, P. Patil and V. Hole, "Vitamins and Minerals
diverse manifestations of vitamin deficiencies. Additionally, Diagnosis System Using Neural Network," 2018 Second International
Conference on Inventive Communication and Computational
it is advisable to explore the creation of a user-friendly mobile
Technologies (ICICCT), Coimbatore, India, 2018, pp. 1921-1925, doi:
application for vitamin deficiency symptom assessment, 10.1109/ICICCT.2018.8473019.
emphasizing simplicity and accessibility. Such an app could [7] J. Ramesh, D. Sankalpa, A. Khamis, A. Sagahyroon and F. Aloul,
"Explainable Machine Learning for Vitamin A Deficiency
empower users to capture and analyze images, contributing Classification in Schoolchildren," 2022 IEEE-EMBS International
to the early detection and intervention of deficiencies. Conference on Biomedical and Health Informatics (BHI), Ioannina,
Moreover, efforts to enhance user accessibility may involve Greece, 2022, pp. 01-04, doi: 10.1109/BHI56158.2022.9926924.
[8] E. K and S. K, "Diagnosis of Vitamin Deficiency in Human Beings
integrating multi-language support, ensuring the application's using DNN Algorithm," 2023 Second International Conference on
usability across diverse linguistic backgrounds. These Electronics and Renewable Systems (ICEARS), Tuticorin, India, 2023,
suggested advancements collectively aspire to enhance the pp. 1627-1632, doi: 10.1109/ICEARS56392.2023.10085334.
[9] R. Guerrero, B. Renteros, R. Castañeda, A. Villanueva and I. Belupú,
model's efficacy and expand its influence. "Detection of nutrient deficiencies in banana plants using deep

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.
 learning," 2021 IEEE International Conference on Automation/XXIV [14] Přemysl Mladěnka, Kateřina Macáková, Lenka Kujovská Krčmová,
Congress of the Chilean Association of Automatic Control (ICA- Lenka Javorská, Kristýna Mrštná, Alejandro Carazo, Michele Protti,
ACCA), Valparaíso, Chile, 2021, pp. 1-7, doi: Fernando Remião, Lucie Nováková, the OEMONOM researchers and
10.1109/ICAACCA51523.2021.9465311. collaborators, Vitamin K – sources, physiological role, kinetics,
[10] V. Krishna, Y. D. Solomon Raju, C. V. Raghavendran, P. Naresh and deficiency, detection, therapeutic use, and toxicity, Nutrition Reviews,
A. Rajesh, "Identification of Nutritional Deficiencies in Crops Using Volume 80, Issue 4, April 2022, Pages 677–
Machine Learning and Image Processing Techniques," 2022 3rd 698, https://doi.org/10.1093/nutrit/nuab061.
International Conference on Intelligent Engineering and Management [15] Z., Guo, X., Zhu, A., He, X., Zhao, X., Han, Y. and Subedi, R. 2020.
(ICIEM), London, United Kingdom, 2022, pp. 925-929, doi: Using deep convolutional neural networks for image-based diagnosis
10.1109/ICIEM54221.2022.9853072. of nutrient deficiencies in rice. Comp. Intell. Neurosci., 20: 1-12.
[11] S. Yadav, A. Rathod, P. Patil and V. Hole, "Vitamins and Minerals https:// doi.org/10.1155/2020/7307252 .
Diagnosis System Using Neural Network," 2018 Second International [16] Velasco Cruz AA, Attié-Castro FA, Fernandes SL, Cortes JF, de Tarso
Conference on Inventive Communication and Computational P Pierre-Filho P, Rocha EM, Marchini JS. Adult blindness secondary
Technologies (ICICCT), Coimbatore, India, 2018, pp. 1921-1925, doi: to vitamin A deficiency is associated with an eating disorder. Nutrition.
10.1109/ICICCT.2018.8473019. 2005 May;21(5):630-3. doi: 10.1016/j.nut.2004.12.003. PMID:
[12] Darnton-Hill I. Public Health Aspects in the Prevention and Control of 15850971.
Vitamin Deficiencies. Curr Dev Nutr. 2019 Jun 21;3(9):nzz075. doi: [17] Investing in the future: A united call to action on vitamin and mineral
10.1093/cdn/nzz075. PMID: 31598578; PMCID: PMC6775441. deficiencies. Global Report 2009
[13] Sommer A. Vitamin deficiency and clinical disease: a historical
overview. J Nutr. 2008 Oct;138(10):1835-9. doi:
10.1093/jn/138.10.1835. PMID: 18806089.

Authorized licensed use limited to: National Institute of Technology Karnataka Surathkal. Downloaded on September 06,2024 at 05:18:07 UTC from IEEE Xplore. Restrictions apply.