Page 86 - sjsi
P. 86
Research Article: Alsibai & Heydari 86
Neural Networks (CNNs) and deep learning rare as well as the ethical issues and sensitivity
have achieved great success in computer vision of such dataset which can pose another
(7)
with its unique advantages . Many diseases problem. Therefore, resorting to a publicly
are diagnosed using different Deep Learning available dataset was chosen to accelerate the
Models (7) . Some examples include the work on this project. After observing some of
detection of COVID-19 using lung ultrasound the related work, one publicly available PCOS
imagery achieving 89.1% accuracy using dataset was utilized in the training of PCONet, a
InceptionV3 network (8) , the use of deep CNN developed by Hosain AK et al. that detects
learning architectures for the segmentation of PCOS from ovarian ultrasound images with
the left ventricle of the heart (9) , and the accuracy of 98.12% on test images as well as
classification of breast lesions in ultrasound fine-tuning InceptionV3 model achieving
images obtaining an accuracy of 90%, sensitivity 96.56% accuracy (11) . The PCOS dataset is
of 86%, and specificity of 96% by utilizing the publicly available on Kaggle (12) . Other related
GoogLeNet CNN (10) . As we can see, deep work includes Wenqi Lv et al. who utilized
learning has proved its potential and the vital image segmentation technique using U-Net on
role it can provide in benefitting and assisting scleral images then a Resnet model was
practitioners that use ultrasonography as a tool adapted for feature extraction of PCOS
for diagnosis. This paper is discussing the achieving classification accuracy of 0.929, and
robustness of deep learning in diagnosing PCOS. AUC of 0.979 (13) . Subrato Bharati et al. used
Since arificial intelligence (AI) and deep learning clinical attributes of 541 women in which 177
algorithms can quickly and reliably assess vast are infected with PCOS to be utilized in a
volumes of data, they can be utilized to machine learning model that uses random
diagnose PCOS in ultrasound scans. AI and deep forest and logistic regression to predict PCOS
learning algorithms are expected to examine disease for which the testing accuracy achieved
ultrasound images to find patterns and traits is 91.01% (14) . Sakshi Srivastava et al. employed
that are indicative of PCOS in the case of PCOS a fine-tuned VGG-16 model to train on their
detection. This can help to increase the speed dataset that consists of ultrasound images of
and accuracy of diagnosis as it can be done the ovaries to detect the presence ovarian cyst
more accurately and efficiently than by manual with 92.11% accuracy obtained (15) . In this
analysis. Furthermore, the application of AI and paper, the Kaggle dataset highlighted
deep learning in the diagnosis of PCOS can previously is used to conduct training by
decrease the workload for medical enhancing the power of transfer learning to
professionals and free them up to concentrate train an existing model architecture that is pre-
on other responsibilities. Overall, the use of AI trained on thousands of images in advance. The
and deep learning in the detection of PCOS in training on this dataset achieves excellent
ultrasound images has the potential to improve results. However, after achieving great results
the accuracy, efficiency, and accessibility of and after further inspection on the dataset, it
healthcare. This was the motive to tackle such turned out that the publicly available PCOS
an important health issue that affects millions dataset that was utilized in training the fine-
of women worldwide and apply the potential of tuned model and that other authors (11) have
deep learning to help nullify this crucial been using in their research is extremely
problem. Obtaining a viable and correct erroneous and full of misleading information.
ultrasound dataset for this task is difficult as the This will be discussed in details in this paper.
annotation of medical images requires
significant professional medical knowledge, Methodology
which makes the annotation very expensive and Dataset Description
SJSI – 2023: VOLUME 1-1
