1. Introduction
A worldwide health concern impacting women of all age groups and ethnic backgrounds
are ovarian cysts. The development of these fluid-filled sacs on or inside the ovaries
can cause a number of health and wellbeing issues for women [1]. The cysts may arise due to a variety of factors, including genetic predispositions,
hormone imbalances, and reproductive diseases. Most of the cysts are benign and go
away on their own, but occasionally they can cause problems like pain, infertility,
and in rare instances, ovarian cancer [2]. There is an immediate requirement to detect and classify ovarian cyst at an early
stage so as to provide proper and effective treatment plan to the patient.
Imaging methods like, magnetic resonance imaging (MRI), computed tomography (CT) and
ultrasound assist healthcare professionals in the diagnosis of ovarian cysts. Ultrasound
imaging technique also known as sonography is the foremost technique suggested by
physicians due to its non-invasive nature, less cost and ability to provide detailed
visualizations of the organs of human body. With high-resolution images, ultrasound
can detect the presence of cysts and find out their size, shape, location and composition.
However, the accuracy of ultrasound imaging is highly based on the skill and expertise
of the radiologist. Inexperienced radiologist may miss subtle signs of ovarian cysts
or misinterpret normal ovarian structures as cysts. In such circumstances, some additional
opinions are required for an accurate and precise decision taken by the physician
[3].
Machine learning (ML) techniques furnish a rule-based system that gives reliable,
automatic and precise decisions [4]. ML algorithms create rules when training is performed on benchmark datasets [5]. These rules are additionally confirmed under the consultation of subject experts.
A proper knowledge representation mechanism is used to organize and store the gained
knowledge in a knowledge base [6]. With the use of these automated procedures, deductive inference may be performed
using this knowledge to help with reasoning and lead to more correct judgments. Deep
learning is subset of machine learning that produces more accurate image analysis
findings. Thus, these models are used to provide accurate, precise and timely analysis
of medical imaging data. The main goal of this paper is to improve the diagnosis and
classification of an ultrasound image as normal ovary or ovary affected with cyst
using deep learning models and thus, offering as a valuable tool for healthcare professionals
in improving medical precision and patient care.
The paper makes the following significant contributions:
• Upon conducting an extensive examination of research articles about the categorization
of ovarian cysts, it is discovered that majority of the studies focused on a limited
number of ultrasound images. This paper provides an ultrasound dataset with a sufficient
number of images to conduct research thoroughly;
• The paper develops four fine-tuned deep learning models to distinguish normal category
ovaries from cyst affected ovaries;
• The proposed model, fine-tuned DenseNet-169 excels in performance and efficiency
against state-of-the-art grounded on evaluation criteria like, accuracy, f1-score,
recall and area under receiver operating curve (AUROC).
The remaining paper is structured as follows: Section 2 deals with the latest deep
learning researches conducted on the ovarian cyst classification. Section 3 encompasses
the ultrasound dataset, methodology, architecture of four fine-tuned deep learning
models. The findings of these models are discussed in section 4 along with the comparison
with previous work. Section 5 concludes the paper and mentions further improvements
that can be incorporated.
2. Related Work
Numerous deep learning algorithms have been developed in recent years for automatic
detection of cysts present in the ovaries using ultrasound images. The authors Raja
and Suresh present a method to uncover and classify various types of ovarian cysts
like, follicular cyst, endometriosis, cystadenoma cyst, cancer cysts, polycystic ovaries,
dermoid cysts, pelvic infections and normal ovaries using a 2D Convolutional Neural
Network model from 240 ultrasound images [7]. Out of 240 images, 160 images are allocated for training purposes and remaining
80 images form the test set. These images are of 128 $\times$ 128 size and are obtained
from different public sectors. The proposed model reaches an accuracy of 99.37% in
detecting and classifying the ovarian cysts.
Another study focuses on the development of convolutional neural networks to differentiate
ovarian endometriosis cyst (OEC) from tubal-ovarian abscess (TOA) [8]. The research involves three Convolutional Neural Networks (CNN) - ResNet-152, DenseNet-161,
and EfficientNet-B7, to contrast between OEC and TOA with the help of 202 ultrasound
images. The models depict high accuracy, sensitivity, and specificity, denoting their
capability as valuable tools for timely screening of these conditions. The results
of these models are compared with assessments by the clinical indicator carbohydrate
antigen 125(CA125) and three ultrasound radiologists. ResNet-152 gives better performance
in distinguishing between OEC and TOA compared to assessments done by radiologists
and CA125 with 0.986 as the area under receiver operating characteristic.
Other researchers introduce a lightweight deep learning model called Ocys-Net to discriminate
several ovarian abnormalities [9]. It uses efficient channel attention module for local cross-channel interaction that
contributes by removing defects generated by dimensionality reduction. This model
extracts shape, textural features from 224 $\times$ 224 ultrasound images. Improved
ShuffleNet V2 is implemented on 750 ultrasound images which has 250 images of normal
ovary, cystic ovary and impure ovarian cyst, respectively. The Ocys-Net gives results
with an accuracy of 95.93%.
The authors [10] present OCD-FCNN (Ovarian Cysts Detection and Classification using Fuzzy Convolutional
Neural Network) for ovarian cysts classification from ultrasound images. The network
first implements a Convolutional Neural Network to unsheathe features from images.
These features are then forwarded as input to Fuzzy Convolutional Neural Network (FCNN)
which is rule based. The FCNN includes fuzzy logic that handles the inherent uncertainty
and ambiguity in classifying ovarian cysts. The proposed OCD-FCNN model evaluates
on a dataset of 440 ovarian ultrasound images from which 320 images are used for training
and 120 images are used for testing. These images are collected from SRM medical science
hospital located at Chennai, TN, India. The dataset comprises of images with normal
ovaries and different type of cysts like, cystadenomas, follicular cysts, dermoid
cysts, endometriosis cysts, pelvic infections, polycystic ovaries and cancer cysts.
The model performs well by achieving an accuracy of 98.37% in the classification of
ovarian cysts.
Harris Hawk Optimization (HHO) with deep Q network is developed by Narmatha and other
researchers [11] to segregate seven categories of ovarian cysts, such as follicular cyst, corpus luteum
cyst, endometriosis cyst, teratoma, dermoid cyst, polycystic ovary and hemorrhagic
cyst. In this method, features of ultrasound images are unsheathed by CNN using Deep
Q network and HHO. This research performs on 478 ultrasound images taken from Royal
Victoria Hospital in Montreal, Canada. Extensive experimental evaluations demonstrate
the effectiveness of HHO-DQN approach to classify ovarian cyst with an accuracy of
97%, f1-measure of 96.5%, precision and recall of 96%. The obtained results exceed
the CNN, artificial neural networks and AlexNet models.
The researchers leverage a hybrid model of deep learning neural network with Support
Vector Machine (SVM) to classify 374 ultrasound images of ovarian cysts [12]. The research aims to develop a classification model that classifies five different
categories of ovarian cysts -- endometriosis cysts, polycystic ovaries, hemorrhagic
cysts, dermoid cysts and malignant cysts. For training and testing of the model, 351
and 23 ultrasound images are used. The model excels from existing classification models
with an accuracy of 97.3% combines deep learning neural networks and Support Vector
Machine to accurately classify ultrasound images of ovarian cysts.
The study aims to develop a neural network-based method for the detection and segmentation
of cysts in ovaries from ultrasound images [13]. Pre-processing methods such as resizing, median filtering and binarization are applied
on ultrasound images to refine the quality of images. Afterwards, region of interest
is drawn out from pre-processed ultrasound images by using bounding box image segmentation
technique. At last, CNN is implemented for extraction of features and classification
of three categories of ovarian cysts - dermoid cysts, endometriosis cysts and hemorrhagic
cysts. The CNN-based approach exhibits promising results (94% accuracy) in accurately
detecting ovaries affected by cysts from 80 ultrasound images.
The research work of Srivastava and other researchers utilizes a fine-tuned VGG-16
deep learning model for classification of ovarian cysts [14]. The research involves fine-tuning the VGG-16 network with 240 ultrasound images
to boost its functioning in detecting ovarian cysts. Out of 240 images, 160 images
are designated for training, while the remaining 80 images are kept for testing. The
endmost four layers of the model are modified to adapt it to the specific task of
ovarian cyst detection. The proposed VGG-16 model denotes a high accuracy rate of
92.11% in detecting ovarian cysts from ultrasound images.
These studies of ovarian cyst classification have uncovered many challenges, such
as preprocessing to enhance ultrasound image quality, classification errors and others.
To improve the classification accuracy of ovarian cysts, the paper aims to develop
a deep learning model with its efficient architecture and feature-rich representations
specifically tailored to detect ovarian cysts in ultrasound images, in turn offering
a reliable and predictive reference for healthcare professionals in the diagnosis
and treatment plan of patients.
3. Materials and Methodology
The methodology proposed in the paper begins with acquiring of ultrasound dataset
of ovaries. The dataset is then pre-processed to increase model's accuracy. After
pre-processing, four CNN models namely, VGG-16 [15], ResNet-152 [16], DenseNet-169 [17] and EfficientNet-B3 [18] are utilized for classification of ultrasound images as either normal ovaries or
cystic ovaries independently. All the models are fine-tuned to achieve better classification.
The model that performs excellent in the evaluation criteria is chosen as the best
model for ovarian cyst classification. The methodology of proposed paradigm to classify
ovarian cyst is shown in Fig. 1.
Fig. 1. Methodology of proposed paradigm to classify ovarian cyst.
3.1. Ultrasound Data Acquisition
The ultrasound dataset employed for this research is taken from ?Skop? medical diagnostic
centre located in Agra, India. These images include transvaginal and abdominal ultrasound
scans and belong to six different categories - follicular cyst, corpus luteum cyst,
endometriosis or chocolate cyst, hemorrhagic cyst, dermoid cyst and normal ovary images.
The images are produced on a Xario 100 (Toshiba, Japan) ultrasound machine which has
the probe frequency of 11MHz. The specifications of ultrasound machine are adjusted
as per need by the radiologist and are the same for the whole dataset. The cases involving
pregnancy, prior ovarian surgery, and poor picture quality are excluded from the research.
In total 1203 ultrasound images of ovaries are used in the research. To protect the
privacy and confidentiality of the patients, the dataset is previously anonymized,
and the images are not connected to any personally identifying information. This large
dataset improves the overall potential and accuracy of the research by offering a
comprehensive outlook on the varied presentations of ovaries affected with cysts.
A sample dataset is displayed in Fig. 2.
Fig. 2. Sample ultrasound dataset.
3.2. Data Pre-processing
Data augmentation is a popular method to handle small datasets [19]. It generates new data points from the pre-existing data and thus, incorporates variations
in the training dataset and also reduces overfitting problem of the model [20]. As a result, it helps in enhancing the model's accuracy with the addition of new
data. In this research, data augmentation is performed by using transformations like,
horizontal flip, vertical flip, random rotation and shear. Thus, each deep learning
model is developed for both original dataset and augmented dataset. For ResNet, DenseNet
and VGG, the input images are resized to 224 $\times$ 224 dimension to fulfil the
models' input requirement and to standardize the distance scale. For EfficientNet-B3,
the expected input size of image is 300 $\times$ 300 and therefore, the input images
are resized to 300 $\times$ 300 to meet the models' input requirement.
3.3. Proposed Deep Learning Models
Convolutional Neural Networks (CNN) are specialized class of deep learning models
that can handle large data of images efficiently with their potential to draw out
hierarchical representations from images [21]. The application areas of CNN revolve around analysis of images, classification,
detection, segmentation and recognition [22-27].
The four CNN models- VGG-16, ResNet-152, EfficientNet-B3 and DenseNet-169 are pretrained
on ImageNet (\href{https://image-net.org/}{https://image-net.org/}), a huge image
database with more than 14 million images. Using ImageNet, training is conducted on
natural images and so model's weights may not work well on medical images. As a result,
transfer learning is used in this research and the pre-trained weights are used for
feature extraction from the ultrasound images.
The dataset is partitioned into two sections, 75% training and 25% testing. Each model
is fine-tuned so as to keep intact the knowledge obtained from pre-training while
allowing the model to change its parameters for better ovarian classification. To
achieve this, instead of using classifying layer of each model, a sequence of a flatten
layer, fully connected layers having 256, 128, 64, 32 and 16 nodes together with interposing
dropout layers and batch normalization layers are used similar to experiments performed
by F. R. Eweje et al. [28]. In the fully connected layers, Rectified Linear Unit function (RELU) activation
function is chosen over different activation functions as it introduces non-linearity
and gradient stability. At the end, the classification layer which has a single node
and sigmoid activation is utilized to accomplish the classification of ovarian cyst
from normal ovaries. The flatten layer converts multidimensional data into a single
dimension but keeping intact all the information needed to connect with fully connected
layers. The batch normalization layer escalates the speed of training by mitigating
the issues related to vanishing gradients.
Step-by-step algorithm:
1. Load necessary libraries and modules.
2. Load the deep learning network that is already pre-trained using ImageNet excluding
top layers.
3. Freeze all layers in the model to avoid being updated during training.
4. Add custom layers - flatten layer, fully connected layers having 256, 128, 64,
32 and 16 nodes with interposing dropout layers and batch normalization layers.
5. Assign a 0.2 dropout value.
6. Review the model summary.
7. Construct the data generator for training.
8. Develop the data generator for testing or validation.
9. Compile and train the model with Adam optimizer, binary cross-entropy loss function
and metrics.
CNN models are executed on Intel(R) Core (TM) i5-5300U CPU and NVIDIA Tesla V100 GPU
using Python (version 3.10) with Keras library. During training, the Adam (Adaptive
Moment Estimation) optimizer with learning rate as 0.01, batch size as 32, drop out
as 0.2 and number of epochs as 300 are set. A threshold value of 0.5 is set to the
last neuron with sigmoid activation function. As this classification is binary, the
loss in training is ascertained using binary cross-entropy loss function. At each
epoch, loss on validation set is tracked and the model with least validation loss
is chosen to depict a training trial. The final classification models are trained
with fine-tuned hyperparameters so as to achieve maximum performance.
3.3.1 VGG model
Visual Geometry Group (VGG) model is a CNN pretrained on ImageNet and developed by
Karen Simonyan and Andrew Zisserman [15] in the University of Oxford in 2014. The input size of image for VGG is 224 $\times$
224. The two versions of VGG models, namely VGG-16 and VGG-19 has 16 and 19 layers,
respectively. The layers comprise of convolutional layers, fully connected layers
and max-pooling layers. An important feature of VGG models is simple architecture
with the use of small size $3 \times 3$ filters. The VGG model secured second position
in the classification process in ImageNet Large Scale Visual Recognition Challenge
(ILSVRC) held in 2014 and thus, demonstrated its capabilities in image analysis tasks.
In this paper, VGG-16 is used and is fine-tuned by replacing the classification layer
with a flatten layer, fully connected layers having 256, 128, 64, 32 and 16 nodes
together along with dropout layers and batch normalization layers.
3.3.2 Residual network model
Residual Networks (ResNet) proposed by Kaiming He et al. in 2015 [16] are based on the concepts of skip connections without affecting performance of the
model. The skip connection bypass one or few layers in between forming residual blocks.
Residual Networks are formed by stacking these residual blocks together. In this paper,
ResNet-152 is employed to classify cystic ovaries from normal ovaries. ResNet-152
is similar to ResNet-50 and ResNet-101 with more depth and complexity. ResNet-152
secured first place in ILSVRC 2015 which sets it as a benchmark in image classification
tasks.
This network uses an input image with a size of 224 $\times$ 224. The architecture
of ResNet-152 consists of 4 stages in which first, second, third and fourth stage
has 3, 8, 36 and 3 residual blocks, respectively. Each residual block consists of
3 convolutional layers, batch normalization layer and ReLU activation function. The
classification layer is replaced with flatten layer, fully connected layers of 256,
128, 64, 32 and 16 nodes accompanied by dropout and batch normalization layers.
3.3.3 Densely connected network model
Densely connected convolutional model is developed by Gao Huang with other researchers
in 2017 [17]. In this network, each layer takes input from all previous layers, increasing the
performance of the network and solving vanishing gradient problem. In this paper,
DenseNet-169 is utilized which takes input of size 224 $\times$ 224 with dense blocks
of 6, 12 and 32 layers interposing with batch normalization, convolutional layer and
average pooling layer. In all convolutional and fully connected layers, ReLU activation
function is used. The classification layer is altered with flatten layer, fully connected
layers of 256, 128, 64, 32 and 16 nodes together with dropout and batch normalization
layers.
3.3.4 EfficientNet mode
EfficientNet is introduced by Tan and Le in 2019 [18]. It is a CNN architecture and scaling method that uses a compound coefficient to
scale depth, breadth, and resolution parameters uniformly. It comprises of Mobile
Inverted Bottleneck (MBConv) layers which are in turn union of depth-wise separable
convolutions and inverted residual blocks. In order to improve accuracy further, it
utilizes Squeeze-and-Excitation (SE) optimization technique. EfficientNet has eight
variants, from EfficientNet-B0 to EfficientNet-B7 with different input shapes ranging
from 224 resolution for EfficientNet-B0 to 600 resolution for EfficientNet-B7. In
this research, EfficientNet-B3 is utilized with resolution of 300 and classification
layer is modified with flatten layer, fully connected layers of 256, 128, 64, 32 and
16 nodes coupled with dropout and batch normalization layers.
3.4. Evaluation Criteria
To assess the performance of the proposed deep learning models, several performance
criteria such as, accuracy, precision, recall, specificity, f1-score and AUROC are
utilized [29]. The mathematical formulae of the evaluation criteria are displayed in Table 1. The parameters used in the formulae are: $TP$ - True Positives, $TN$ - True Negatives,
$FP$ - False Positives and $FN$ - False Negatives.
Table 1. Evaluation criteria.
|
Parameters
|
Formulae
|
|
Accuracy
|
$\frac{TP+TN}{FP+FN\mathrm{+\ }TP+TN}$
|
|
Recall
|
$\frac{TP}{FN+TP}$
|
|
Precision
|
$\frac{TP}{FP+TP}$
|
|
F1-score
|
$\frac{\mathrm{2*}Recall*Precision}{Recall+Precision}$
|
|
Specificity
|
$\frac{TN}{FP+TN}$
|
4. Results and Discussion
This section talks about the capabilities of four fine-tuned deep learning models
to classify ovarian cyst. In this research, a medical dataset consisting of 1203 ultrasound
images of ovaries is used to train VGG-16, ResNet-152, EfficientNet-B3 and DenseNet-169
models. To further expand the dataset and add more variations among the images, data
augmentation is used. Thus, two sets of datasets - an original dataset and augmented
dataset are used for training and evaluation of proposed deep learning models.
The training accuracy and loss along with validation accuracy and loss of these four
models for 300 epochs on the original dataset and augmented dataset are shown in Figs. 3 and 4, respectively. In few instances, accuracy has reached up to 100 percent. It is observed
that accuracy obtained using augmented dataset is greater than that of the original
dataset. Thus, data augmentation definitely helps in enhancing the performance of
a deep learning model.
Fig. 3. Training accuracy and loss curves using original dataset.
Fig. 4. Training accuracy and loss curves using augmented dataset.
Table 2 presents the evaluation criteria for the four fine-tuned CNN models used in the research
on original dataset. VGG-16 model has obtained accuracy of 88.23%, precision of 82%,
highest recall of 100%, f1-score of 90.11%, specificity of 74.65%. ResNet-152 and
EfficientNet-B3 has achieved average performance criteria between 65% to 83%. DenseNet-169
has performed well in all criteria with the values of 92.81% for accuracy, 89.01%
for precision, 93.64% for f1-score, 85.92% for specificity and 98.78% for recall.
These results are emphasized in the table. Fig. 5 show the comparative analysis of VGG-16, ResNet-159, EfficientNet-B3 and DenseNet-169
using original dataset. Fig. 6 outlines the AUROC curves for the four CNN models, in which DenseNet-169 gets the
maximum value of 0.9235 and ResNet-152 has the minimum value of 0.7180.
Table 2. Evaluation criteria on original dataset.
|
Criteria
|
Fine-tuned VGG-16
|
Fine-tuned ResNet-152
|
Fine-tuned EfficientNet-B3
|
Fine-tuned DenseNet-169
|
|
Accuracy
|
88.23
|
71.89
|
74.51
|
92.81
|
|
Recall
|
100
|
73.17
|
82.93
|
98.78
|
|
Precision
|
82.0
|
74.07
|
73.12
|
89.01
|
|
F1-score
|
90.11
|
73.62
|
77.71
|
93.64
|
|
Specificity
|
74.65
|
70.42
|
64.79
|
85.92
|
Fig. 5. Comparative analysis of evaluation criteria using original dataset.
Fig. 6. AUROC curve using original dataset.
In the case of augmented dataset, the evaluation criteria are displayed in Table 3. The performance of the four fine-tuned models, VGG-16, ResNet-152, EfficientNet-B3
and DenseNet-169 are inspected on the testing results with accuracies of 99.34%, 83.07%,
97.14% and 99.78%, respectively. DenseNet-169 has obtained maximum accuracy of 99.78%
as well as other performance measures like, recall as 100%, f1-score as 99.79%, precision
as 99.58% and specificity as 99.54%. ResNet-152 has achieved minimum accuracy of 83.07%
with other measures like precision, recall, f1-score, specificity of 93.89%, 71.91%,
81.44% and 95%, respectively. EfficientNet-B3 has achieved 100% in precision and specificity
measures. It has also attained 97.14% in accuracy, 94.47% in recall and 97.15% in
f1-score.
Table 3. Evaluation criteria on augmented dataset.
|
Criteria
|
Fine-tuned VGG-16
|
Fine-tuned ResNet-152
|
Fine-tuned EfficientNet-B3
|
Fine-tuned DenseNet-169
|
|
Accuracy
|
99.34
|
83.07
|
97.14
|
99.78
|
|
Recall
|
99.57
|
71.91
|
94.47
|
100
|
|
Precision
|
99.15
|
93.89
|
100
|
99.58
|
|
F1-score
|
99.36
|
81.44
|
97.15
|
99.79
|
|
Specificity
|
99.09
|
95.0
|
100
|
99.54
|
For augmented dataset, Fig. 7 outlines how well DenseNet-169 performs when compared to VGG-16, ResNet-152 and EfficientNet-B3
in terms of accuracy. The graph shows that DenseNet-169 model is 99.78% accurate
and hence can be used as a tool to classify cystic ovaries from normal ovaries. The
graphical representation of other evaluation measures such as precision, recall, f1-score
and specificity is depicted in Fig. 8. Fig. 9 illustrates AUROC for the four fine-tuned deep learning models. The fine-tuned DenseNet-169
has achieved the largest AUROC as 0.9977. The result analysis proves that DenseNet-169
has performed well compared to VGG-16, ResNet-152 and EfficientNet-B3 and can be used
as an effective method for ovarian cyst classification. This again proves that data
augmentation plays an important part in improving the performance of deep learning
models.
Fig. 7. Accuracy curve analysis using augmented dataset.
Fig. 8. Comparative analysis of evaluation criteria using augmented dataset.
Fig. 9. AUROC curve using augmented dataset.
Table 4 shows the evaluation criteria for the four CNN base models used in the research on
augmented dataset. From the table, it is quite evident that DenseNet-169 base model
has performed better in accuracy, recall and f1-score compared to other base models.
It has achieved 97.89% in accuracy, 99.20% in recall, 98.53% in precision, 97.91%
in f1-score and 98.14% in specificity. The comparative analysis of Tables 3 and 4
gives an indication that fine-tuning has enhanced the performance of all the deep
learning models used in the research. And fine-tuned DenseNet-169 is the best performer
deep learning model for classification of ovarian cyst.
Table 4. Evaluation criteria of base models on augmented dataset.
|
Criteria
|
VGG-16
|
ResNet-152
|
EfficientNet-B3
|
DenseNet-169
|
|
Accuracy
|
97.45
|
81.97
|
95.94
|
97.89
|
|
Recall
|
98.56
|
69.98
|
93.34
|
99.20
|
|
Precision
|
98.05
|
92.34
|
98.92
|
98.53
|
|
F1-score
|
97.83
|
80.12
|
96.85
|
97.91
|
|
Specificity
|
98.79
|
93.64
|
99.45
|
98.14
|
The research also compares the Mean Squared Error (MSE) and cross entropy loss of
all fine-tuned deep learning models on augmented dataset in Table 5. DensetNet-169 has minimum MSE (0.0022) and cross entropy loss (0.0074) compared
to corresponding values of VGG-16, ResNet-152 and EfficientNet-B3. Henceforth, the
fine-tuned DenseNet-169 has proven to be a better classification model for ovarian
cyst classification.
Table 5. MSE and cross entropy loss on augmented dataset.
|
|
Fine-tuned VGG-16
|
Fine-tuned ResNet-152
|
Fine-tuned EfficientNet-B3
|
Fine-tuned DenseNet-169
|
|
MSE
|
0.0065
|
0.1692
|
0.0286
|
0.0022
|
|
Cross Entropy Loss
|
0.01053
|
0.04937
|
0.4304
|
0.0074
|
This research further compares the proposed fine-tuned DenseNet-169 model results
with the latest ovarian cyst classification models mentioned in the literature that
considered implementations on 75-25% training and testing data split. A summary of
this contrast is shown in Table 6. Deep Q Network with Harris Hawk Optimization outperformed previous works with an
accuracy of 97%. With further improvements in deep learning models, Ravishankar along
with other researchers developed a Fuzzy based CNN classifier model with an accuracy
of 98.37%.
Table 6. Comparison between findings of proposed model with recent deep learning models.
|
References
|
Year
|
Models
|
Accuracy (%)
|
|
[8]
|
2023
|
ResNet-152, DenseNet-161, EfficientNet-B7
|
96.80
|
|
[9]
|
2023
|
Improved ShuffleNet V2
|
95.93
|
|
[10]
|
2023
|
Fuzzy Convolutional Neural Network (FCNN)
|
98.37
|
|
[11]
|
2023
|
Deep Q Network with Harris Hawk Optimization
|
97.0
|
|
[13]
|
2022
|
CNN
|
94.0
|
|
Proposed
|
2024
|
Fine-tuned DenseNet-169
|
99.78
|
The fine-tuned DenseNet-169 model proposed in this research outshines all the previous
deep learning models and gives a better evaluation boost in relation to classification
accuracy of 99.78%. Consequently, it can be utilized as a tool for healthcare professionals
in improving predictable accuracy and patient care related to ovarian pathologies.
5. Conclusion and Future Scope
An accurate and efficient deep learning ovarian cyst classification model with minimal
data pre-processing is proposed in this research. It employs fine-tuned VGG-16, ResNet-152,
EfficientNet-B3 and DenseNet-169. Several evaluation criteria like, accuracy, recall,
precision, specificity, f1-score and AUROC are utilized to check their robustness.
It is noteworthy that the fine-tuned DenseNet-169 model demonstrates maximum classification
accuracy of 99.78%. It has also surpassed several classification models mentioned
in the latest literature of ovarian cyst on similar datasets. Henceforth, it can be
deployed as decision support tool to assist radiologists and health care professionals
in interpreting imaging studies and making treatment recommendations, ultimately improving
patient outcomes in the diagnosis and management of ovarian cysts.
The findings of this research can be extended to categorize several kinds of ovarian
cysts. It can also be generalized to other clinical settings and accurately identify
disorders of a similar nature using ultrasound, CT or MRI images.
In summary, we have designed a deep learning model that accurately classifies ovarian
cysts from normal ovaries. Thus, this research is a definite help to healthcare professionals
for giving patients with appropriate and timely clinical diagnoses of ovarian cysts.
ACKNOWLEDGMENTS
The authors would like to express their gratitude towards Dr. Pallavi Goyal for
sharing a set of anonymised ultrasound images and for her insight advice regarding
how to interpret these images medically.
REFERENCES
A. M. Parekh and N. B. Shah, ``Classification of ovarian cyst using soft computing
technique,'' Proc. of 8th International Conference on Computing, Communications and
Networking Technologies, ICCCNT 2017, no. September, 2017.
S. Mobeen and R. Apostol, ``Ovarian cyst,'' National Library of Medicine, 2022. https://www.ncbi.nlm.nih.gov/books/NBK560541/(accessed
Nov. 20, 2023).
A. A. Alzubi, H. Al Moghrabi, M. A. Alzubi, S. A. Alzubi, and S. Arabia, ``Methods
for Automatic Cyst Detection and Classification in Ultrasound Images of the Female
Genitalia Using Image Processing,'' Journal of Population Therapeutics and Clinical
Pharmacology, vol. 30, no. 6, pp. 297-305, 2023.
B. Mahesh, ``Machine learning algorithms - A review,'' International Journal of Science
and Research, vol. 9, no. 1, January 2020, pp. 381-386, 2020.
A. L. Tarca, V. J. Carey, X. Chen, R. Romero, and S. Drǎghici, ``Machine learning
and its applications to biology.,'' PLoS Computational Biology, vol. 3, no. 6. 2007.
J. Alzubi, A. Nayyar, and A. Kumar, ``Machine learning from theory to algorithms,''
Journal of Physics: Conference Series, 2018.
P. Raja and P. Suresh, ``Variety of ovarian cysts detection and classification using
2D convolutional neural network,'' Multimedia Tools and Applications, vol. 83, no.
16, pp. 49473-49491, 2024.
P. Hu, Y. Gao, Y. Zhang, and K. Sun, ``Ultrasound image-based deep learning to differentiate
tubal-ovarian abscess from ovarian endometriosis cyst,'' Frontiers in Physiology,
vol. 14, pp. 1-9, 2023.
J. Fan, J. Liu, Q. Chen, W. Wang, and Y. Wu, ``Accurate ovarian cyst classification
with a lightweight deep learning model for ultrasound images,'' IEEE Access, vol.
11, pp. 110681-110691, 2023.
T. N. Ravishankar, H. Makarand Jadhav, N. Satheesh Kumar, S. Ambala, and M. Pillai
N, ``A deep learning approach for ovarian cysts detection and classification (OCD-FCNN)
using fuzzy convolutional neural network,'' Measurement: Sensors, vol. 27, no. May,
100797, 2023.
C. Narmatha, P. Manimegalai, J. Krishnadass, P. Valsalan, S. Manimurugan, and M. Mustafa,
``Ovarian cysts classification using novel deep reinforcement learning with Harris
Hawks optimization method,'' Journal of Supercomputing, vol. 79, no. 2, pp. 1374-1397,
2023.
Y. Suganya, S. Ganesan, and P. Valarmathi, ``Ultrasound ovary cyst image classification
with deep learning neural network with Support vector machine,'' International Journal
of Health Sciences, vol. 6, pp. 8811-8818, 2022.
R. Benazir Begam, V. Yogalakshmi, G. Saranya, D. Gururaj, S. Jagtap, and V. Ravanan,
``Ovarian cyst detection using neural networks,'' Proc. of the International Conference
on Electronics and Renewable Systems, ICEARS 2022, pp. 1827-1830, 2022.
S. Srivastava, P. Kumar, V. Chaudhry, and A. Singh, ``Detection of ovarian cyst in
ultrasound images using fine-tuned VGG-16 deep learning network,'' SN Computer Science,
vol. 1, no. 2, pp. 1-8, 2020.
K. Simonyan and A. Zisserman, ``Very deep convolutional networks for large-scale image
recognition,'' Proc. of International Conference on Learning Representations, pp.
1-14, 2014.
K. He, X. Zhang, S. Ren, and J. Sun, ``Deep residual learning for image recognition,''
Proc. of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,
pp. 770-778, 2016.
G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, ``Densely connected convolutional
networks,'' Proc. of 30th IEEE Conference on Computer Vision and Pattern Recognition,
CVPR 2017, pp. 2261-2269, 2017.
M. Tan and Q. V. Le, ``EfficientNet: Rethinking model scaling for convolutional neural
networks,'' Proc. of 36th International Conference on Machine Learning, ICML 2019,
pp. 10691-10700, 2019.
Y. Wang, G. Huang, S. Song, X. Pan, Y. Xia, and C. Wu, ``Regularizing deep networks
with semantic data augmentation,'' IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 44, no. 7, pp. 3733-3748, 2022.
B. Kayalibay, G. Jensen, and P. van der Smagt, ``CNN-based segmentation of medical
imaging data,'' arXiv preprint arXv:1701.03056, 2017
G. Litjen, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompo, M. Ghafoorian, J. A.
W. M. van der Laak, B. van Ginneken, and C. I. Sánchez, “A survey on deep
learning in medical image analysis,” Medical Image Analysis, vol. 42, pp. 60-88,
2017.
S. E. Nassar, I. Yasser, H. M. Amer, and M. A. Mohamed, ``A robust MRI-based brain
tumor classification via a hybrid deep learning technique,'' Journal of Supercomputing,
vol. 80, no. 2, pp. 2403-2427, 2024.
A. Kodipalli, S. L. Fernandes, and S. Dasar, ``An empirical evaluation of a novel
ensemble deep neural network model and explainable AI for accurate segmentation and
classification of ovarian tumors using CT images,'' Diagnostics, vol. 14, no. 5, 2024.
A. Kodipalli, S. L. Fernandes, V. Gururaj, S. V. Rameshbabu, and S. Dasar, ``Performance
analysis of segmentation and classification of CT-scanned ovarian tumours using U-Net
and deep convolutional neural networks,'' Diagnostics, vol. 13, no. 13, 2023.
R. V. Manjunath, A. Ghanshala, and K. Kwadiki, ``Deep learning algorithm performance
evaluation in detection and classification of liver disease using CT images,'' Multimedia
Tools and Applications, vol. 83, no. 1, pp. 2773-2790, 2024.
C. Kamala and J. M. Shivaram, ``Segmentation of ovarian cyst using improved U-NET
and hybrid deep learning model,'' Multimedia Tools and Applications, vol. 83, no.
14, pp. 42645-42679, 2024.
A. Bruno, G. Ignesti, O. Salvetti, D. Moroni, and M. Martinelli, ``Efficient lung
ultrasound classification,'' Bioengineering, vol. 10, no. 5, 2023.
F. R. Eweje, B. Bao, J. Wu, D. Dalal, W. Liao, Y. He, Y. Luo, S. Lu, P. Zhang, X.
Peng, R. Sebro, H. X. Bai, and L. States, ``Deep Learning for Classification of Bone
Lesions on Routine MRI,'' EBioMedicine, vol. 68, 2021, doi: 10.1016/j.ebiom.2021.103402.
M. Kalaiyarasi, R. Dhanasekar, S. Sakthiya Ram, and P. Vaishnavi, ``Classification
of benign or malignant tumor using machine learning,'' IOP Conference Series: Materials
Science and Engineering, vol. 995, no. 1, 2020.
Author
Aditi Gupta an Associate Professor in the Department of Computer Science & Engineering,
Anand School of Engineering & Technology, Sharda University, Agra, India She received
her B.E. degree in Computer engineering from North Maharashtra University, Jalgaon,
India. She did her M.Tech. degree in computer science and engineering from MNNIT,
Allahabad, India. She has 10 years of teaching experience and 3 years of corporate
work experience. Her areas of interest include machine learning, deep learning, natural
language processing, text analytics and medical images.
Hoor Fatima is an Assistant Professor in the School of Computer Science Engineering
and Technology, Bennett University, Greater Noida, UP, India. She holds a Ph.D. degree
from the Indian Institute of Technology, Delhi. She has completed her B.Tech. degree
in computer science and engineering from UPTU. She has published several papers in
the Scopus indexed journals and google scholar citations. Her areas of interest are
machine learning, artificial intelligence, soft computing, deep learning, solar energy
and renewable energy.