Ultrasound imaging processing technology has been employed for more than 50 years. Although it has developed quickly, it has some advantages and particular challenges. It is crucial to establish the fetal survival rate, gestational age, and other factors early on, from the standpoint of ultrasound picture analysis. In a bid to better understand the fetus's continuing growth, fetal anatomy ultrasound image analysis techniques have recently been studied and have emerged as an essential tool for prenatal anomaly diagnosis. The moment has come to thoroughly analyse prior efforts in this area and forecast future directions. Thus, this article discusses cutting-edge methods along with fundamental concepts, theories, and advantages and disadvantages of ultrasound picture technology for the entire fetal along with different anatomies. It begins by summarizing the ongoing issues and introducing the widely used image processing techniques, such as classification, segmentation, etc. We provided the necessary data performance on table 1 that is Fig. 4.
Research Article
Ultrasound 2d Fetal Developing Brain Image Classification and Disease Prediction
https://doi.org/10.21203/rs.3.rs-2748807/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
Fetal
Googlenet
Image classification
Ultrasound
The first use of ultrasound (US) for brain surgery was reported by Chandler et al., describing the surgical results of 21 cases using two-dimensional imaging (2D-US), allowing real-time visualization of the underlying anatomy and pathology throughout the pathology performance. Since then, without being exposed to ionizing radiation, the use of intraoperative ultrasound has allowed surgeons to make better decisions during a surgical procedure. Smaller probes and more seamless integration with neuron navigation systems are examples of how the technology has advanced along with the advancements in neuroimaging modalities and image quality. The development of related developments is another in the context of these advances, we use the 2D-US in comparison to other modalities for fetal brain development. To extract pertinent information from the photos and categorize the fetal into normal or abnormal categories, this field combines sophisticated image processing techniques and machine learning algorithms. Additionally, certain illnesses including cerebral palsy, microcephaly, and Down syndrome can be predicted using the algorithms.
Figure.1 displays actual samples of fetal brain images. This method has the potential to significantly increase the accuracy and early diagnosis of fetal brain abnormalities, enabling earlier interventions and treatments that can significantly enhance the result both for the mother and the fetus. However, as this is a new subject, more study is required to enhance the precision and dependability of the techniques employed in the classifying and prognostication of diseases from unborn baby brain images.Because of the increased prevalence of genetic problems in newborns, the application of ultrasound 2D fetal developing brain picture categorization and illness prediction is becoming more and more crucial. The prognosis for the afflicted children and their families can be considerably improved by early detection and treatment of these diseases.
In this discipline, several machine learning techniques are utilized to evaluate ultrasound pictures and categorize the fetus into normal or pathological categories. These algorithms include like deep-neural networks and decision trees, and lastly support-vector machines. Additionally, the likelihood of certain illnesses may be predicted by algorithm. Despite the potential advantages of this strategy, further study is necessary to increase the precision and dependability of the techniques employed because the area is still in its infancy. The lack of high-quality ultrasound pictures for use in algorithm testing and training is one of the difficulties in this subject. The variation in the pictures brought on by the various acquisition methods and the location of the fetus inside the uterus presents another difficulty.
Considering these difficulties, the topic of ultrasound 2D fetal developing brain picture categorization and illness prediction has significant promise for enhancing fetal brain problem detection and therapy. Healthcare providers may improve care for pregnant women and their fetuses by developing technologies and improving the algorithms they employ.
Recent developments include the categorization and illness prediction of Fetal brain ultrasound imaging in two dimensions. Implementing methods for deep learning such as convolutional neural networks, is one such approach (CNNs), to increase the precision of illness prediction and fetal brain picture categorization.
The time and effort needed by medical professionals to manually examine the pictures would be reduced if automated solutions for ultrasound 2D fetal brain image processing were developed. Additionally, this would enhance the consistency and precision of the analysis, improving patient outcomes.
The ethical and legal issues associated with the categorization and illness prediction of ultrasound 2D fetal brain images are also significant. This covers concerns about informed consent, confidentiality, and data protection, as well as the precision and dependability of the forecasts produced by these methods.
Additionally, more study is required to confirm the precision and dependability of these methods and to ascertain any potential long-term consequences of ultrasound radiation exposure on the growing fetus. To guarantee consistency and comparability of results across various studies and populations, additional standardization in the acquisition and processing of ultrasound images is also required.
Overall, the field of ultrasound 2D fetal brain image classification and disease prediction holds great promise for enhancing mother and baby health and well-being, but it is crucial to take into account and address the various ethical, legal, and practical issues associated with its creation and application.
The descriptiveness and discrimination capability of de- rived features are essential for achieving effective analysis performance in image analysis tasks. Because the features for recognition may be automatically extracted by training, deep learning has the benefit that this sort of method can be extended to difficult situations with very complicated characteristics.
A. CONVOLUTIONAL NEURAL NETWORK
Artificial neural networks known as deep neural networks (ConvNets or CNNs) are used for natural language processing as well as image and video recognition. They are made to handle data having a grid-like architecture, such as an image while maintaining the spatial link between the pixels by employing convolutional layers to learn local characteristics.
A ConvNet is composed of an input layer, hidden convolutional layers, pooling layers, fully connected layers, and output layers. The convolutional layers, which apply filters to the incoming data, create the feature map. The pooling layer reduces the spatial size of the feature map, the number of variables in the network, and allows for the detection of features of varying sizes. Figure 2 explains the process of a Convolutional Neural Network.
B. GOOGLE-NET
GoogleNet and other The use of Convolutional Neural Networks (CNNs) in analysis of ultrasound images in fetal imaging to improve the accuracy of fetal diagnostics. CNNs have shown promising results in a number of fetal imaging applications, including fetal growth estimation, fetal biometry measurement, and fetal anomaly detection.
In fetal ultrasound imaging, CNNs can be mainly used to the – automation of the extract features from the ultrasound images and make predictions about various aspects of the fetus, such as gestational age, fetal weight, and the presence of anomalies. These predictions can then be used to support or improve clinical decision-making.
One of the major benefits of utilizing CNNs for fetal ultrasound analysis is was their - ability to learn from large amounts of data, which can improve the accuracy of their predictions. In addition, they can be trained end-to-end, which means that they can learn to make predictions directly from raw ultrasound images, without the necessity for segmentation algorithm or feature extraction.
Overall, the use of CNNs in fetal ultrasound imaging has the potential to improve the accuracy of fetal diagnostics and make them more accessible to a wider range of healthcare providers.
It's important to keep in mind that the application of CNNs in fetal ultrasound imaging is still an emerging field, and further research is needed to fully evaluate their performance and assess their impact on clinical practice. However, the potential benefits of using CNNs in fetal ultrasound imaging are significant and demonstrate the potential for deep learning to transform healthcare and improve patient outcomes.
Figure 3 clearly explains In the modern world, GoogleNet is utilized for a variety of computer vision applications, that is Object Detection as well as Image Classification.
C. DISCRETEWAVELET TRANSFORM (DWT)
The image quality and reduce noise. The process works by decomposing the image into different frequency components using wavelets, which are mathematical functions that help to analyse and represent signals. The high-frequency components of the image, which contain most of the noise, are then suppressed or thresholder to reduce their impact on the image. The resulting wavelet coefficients are then inverse-transformed back into the image domain to produce a denoised and improved version of the original image. This process can enhance the visibility of fine structures and improve the diagnostic accuracy of the ultrasound examination.
Table 1 COMPARISON OF DATA PERFORMANCE
The above table shows Fig. 4 each stage of the comparison of DWT and CNN.
-
Image acquisition: The original image is acquired using a 2D ultrasound machine.
-
Decomposition: The image is decomposed using a wavelet transform, resulting in a set of wavelet coefficients that represent the image's numerous frequency components.
-
Thresholding: The In the context of 2D fetal ultrasound, DWT (Discrete Wavelet Transform) is used to improve high-frequency coefficients, which contain most of the noise, the threshold to reduce their impact on the image. This can be done using various thresholding methods, such as hard thresholding or soft thresholding.
-
Reconstruction: The threshold coefficients are inverse-transformed back to the image domain-produce a denoised and updated version of the real image.
-
Image display: The resulting image is displayed on the ultrasound machine's screen for interpretation by a trained medical professional.
The Fig. 5 clearly shows the difference between the progress of K Means, DWT, and CNN.
For the DWT wavelet, the wavelets are sampled at regular intervals. DWT provides data about both the spatial and sensitive attributes of a picture at the same time. To evaluate an image, the Discrete wavelet transform method can combine the analyzing filters bank and decimate operation.
Each decomposition level's low and high pass filters are included in the analysis filter bank. While a less-level band pulls the necessary details from data, this same higher-level gathers elements like edges. Two distinct 1D transforms are used to create the 2D transform. In a 1-dimensional Discrete wavelet transform, the approximate coefficients pattern frequency components whereas the detailed components convey higher frequency components.
The above graph that if Fig. 6 clearly explain the graph level of (DWT, CNN, K-Means)
The input signal is split up into four separate subsets when 2-dimensional DWT is used: lower frequencies elements as in longitudinal and transverse directions (cA), lower frequencies elements inside the longitudinal and high-frequencies elements from the parallel bars (cV), higher-frequencies elements with the longitudinal and minimal wavelet coefficients of the parallel bars (cH), and higher - frequencies elements in the longitudinal then the transverse direction. Usually referred to as cA, cV, cH, and cD, respectively. A reconstruction of the following 1-level Discrete wavelet transform and Using Eq. 1, the post arises provided.
I = I1a+ I1h+I1v+I1d (1)
Where I1 h, I1 v, and I1 d stand for horizontal, vertical, and diagonal features, respectively, and I1 is a representation of the input image's closest estimate. The strengths of the words reveal the level of breakdown. By gradually decomposing the LL subband, additional decompositions can be made, and the resulting image can then be divided into several bands. Eq. 2 represents a picture after 5-level DWT decomposition.
I = I5a + 5∑i=1{I1h1 + Iv1+Id1} (2)
We used the reverse biorthogonal family of wavelets as well as wavelets in one and two dimensions for our paper. By implementing edge-tracked scale normalization before the DWT procedure, effective feature extraction was accomplished. The scaled basis function is used by the biorthogonal and reverse biorthogonal wavelets in order to decompose and rebuild an image from one resolution level to the next.
The converted data can be sorted with a resolution that is appropriate for its scale thanks to the usage of DWT as a feature extractor. Small and large characteristics can both be seen since they may be investigated individually thanks to the converted image's multi-level representation.
Since DWT are not similar or match to the Trigonometric function transform, DWT handles data discontinuities better than Discrete Cosine Transform (DCT). As an outcome, DWT is a powerful decoder for complex data such as Color FERT and cmu pie, resulting in higher results
The above block diagram which is Fig. 7 shows the actual process of how the images will be trained and classified.
The data flow in the use of CNNs for fetal ultrasound imaging can be described as follows:
-
Input image: The input to the system is an ultrasound image of a fetus, typically acquired using a transabdominal or transvaginal probe. The image may need to be pre-processed to correct for artifacts, such as speckle noise, and to enhance the picture quality.
-
Training: The data are given as source and they may undergo various pre-process steps, such as cropping, resizing, and normalization, to standardize the size and format of the images and improve the performance of the CNN.
-
Feature extraction: The CNN then extracts features from the pre-processed image. This involves a series of convolutional, pooling, and normalization operations that are designed to identify and isolate relevant features in the image.
-
Apply algorithm: The extracted features are then used to make predictions about various aspects of the fetus, such as gestational age, fetal weight, or the presence of anomalies. This is done by applying an algorithm, such as a multi-layer perceptron or a support vector machine, to the extracted features.
The Fig. 8 shows the distinctive architecture of GoogleNet, which consists of a number of inception modules, is well recognized. A sophisticated building piece called an inception module is made up of several branches, each with a unique arrangement of filters and kernel sizes.
By doing this, the network may extract many complementing characteristics from the input image.
Initially, the images are pre-loaded and the data are saved and then it is pre-processed. The images are selected from the database and then the images are started to train. Discrete wavelet transformation is used to make the image more accurate and it changes the low-level and unclear images into clear images. It helps the user to find the normal/abnormalities in that image. Then by using the convolution neural network algorithm google net, the images are scanned and found whether it is normal or abnormal. If it results in abnormality, it shows the details of that disease and what precautions can be done.
The classification of 2D fetal developing brain pictures and the prediction of linked disorders using CNNs can yield promising results. Anomalies like ventriculomegaly and encephalocele, as well as other anomalies, can be accurately detected in fetal brain scans using CNNs, according to several studies. With some research claiming accuracy rates of over 90%, these results have demonstrated that CNNs can detect these anomalies with a high degree of consistency and precision.
Additionally, by automating laborious and arbitrary processes like manual feature extraction and segmentation, CNNs can increase the effectiveness of fetal brain diagnosis. In particular, in situations with low resources, this can lessen the workload of healthcare professionals and increase the accessibility of prenatal brain diagnostics.
Initially, the images are pre-loaded and the data are saved and then it is pre-processed. The images are selected from the database and then the images are started to train. Discrete wavelet transformation is used to make the image more accurate and it changes the low-level and unclear images into clear images. It helps the user to find the normal/abnormalities in that image. Then by using the convolution neural network algorithm google net, the images are scanned and found whether it is normal or abnormal. If it results in abnormality, it shows the details of that disease and what precautions can be done.
That's really crucial to remember that a variety of elements, including the caliber of the input images, the choice of CNN architecture, and the availability of annotated training data, might influence the outcomes of utilizing CNNs for fetal brain imaging categorization and disease prediction. As a result, it's crucial to thoroughly assess each study's findings and to take these things into account when interpreting them.
Figure 8 is nothing but it is the sample output of the brain image classification.
Overall, the results of utilizing CNNs to classify 2D images of the developing brain in fetal and forecast disorders associated with them are encouraging, and also have the ability and knowledge of machine learning to increase the precision and accessibility of fetal brain diagnostics. To thoroughly evaluate their effectiveness and determine their impact on clinical practice, more research is necessary.
In conclusion, the use of CNNs for the classification of 2D fetal developing brain images and the prediction of related diseases has the potential to transform fetal brain diagnostics and improve patient outcomes, but further research and development are needed to fully realize its potential.
Conflict of Interest: The authors declare that they have no conflict of interest.
Author contributions: Gokul N designed the study, collected and analyzed the data, and wrote the manuscript. Sridharan K contributed to the study design, assisted with data analysis. Giridharan L helped with data collection and interpretation and provided editorial support for the manuscript. Akashdeep V assisted with data analysis and provided statistical expertise. All authors have read and approved the final version of the manuscript.
Funding: This research did not receive any external funding.
- American College of Radiology, “ACR-SPR practice parameter for the safe and optimal performance of fetal magnetic resonance imaging (MRI), resolution 45,” https://www.acr.org/-/media/ACR/Files/ Practice-Parameters/mr-fetal.pdf, 2020.
- C. L. Herrera, J. J. Byrne, H. R. Clark, D. M. Twickler, and J. S. Dashe, “Use of fetal magnetic resonance imaging after sonographic identification of major structural anomalies,” Journal of Ultrasound in Medicine, vol. 39, no. 10, pp. 2053–2058, Oct. 2020.
- A. Gholipour, J. A. Estroff, and S. K. Warfield, “Robust super-resolution volume reconstruction from slice acquisitions: Application to fetal brain MRI,” IEEE Transactions on Medical Imaging, vol. 29, no. 10, pp. 1739–1758, Oct. 2010.
- M. Kuklisova-Murgasova, G. Quaghebeur, M. A. Rutherford, J. V. Hajnal, and J. A. Schnabel, “Reconstruction of fetal brain MRI with intensity matching and complete outlier removal,” Medical Image Analysis, vol. 16, no. 8, pp. 1550–1564, Dec. 2012.
- B. Hou, B. Khanal, A. Alansary, S. McDonagh, A. Davidson, M. Rutherford, J. V. Hajnal, D. Rueckert, B. Glocker, and B. Kainz, “3-D reconstruction in canonical co-ordinate space from arbitrarily oriented 2-D images,” IEEE Transactions on Medical Imaging, vol. 37, no. 8, pp. 1737–1750, Aug. 2018.
- S. S. M. Salehi, S. Khan, D. Erdogmus, and A. Gholipour, “Real-time deep pose estimation with geodesic loss for image-to-template rigid registration,” IEEE Transactions on Medical Imaging, vol. 38, no. 2, pp. 470–81, Feb. 2019.
- M. Ebner, G. Wang, W. Li, M. Aertsen, P. A. Patel, R. Aughwane, A. Melbourne, T. Doel, S. Dymarkowski, P. D. Coppi, A. L. David, J. Deprest, S. Ourselin, and T. Vercauteren, “An automated framework for localization, segmentation and super-resolution reconstruction of fetal brain MRI,” NeuroImage, vol. 206, no. 116324, Feb. 2020.
- A. Singh, S. S. M. Salehi, and A. Gholipour, “Deep predictive motion tracking in magnetic resonance imaging: Application to fetal imaging,” IEEE Transactions on Medical Imaging, vol. 39, no. 11, pp. 3523–3534, Nov. 2020.
- E. Ferrante and N. Paragios, “Slice-to-volume medical image registration: A survey,” Medical Image Analysis, vol. 39, pp. 101–123, Jul. 2017.
- A. Uus, T. Zhang, L. H. Jackson, T. A. Roberts, M. A. Rutherford, J. V. Hajnal, and M. Deprez, “Deformable slice-to-volume registration for motion correction of fetal body and placenta MRI,” IEEE Transactions on Medical Imaging, vol. 39, no. 9, pp. 2750–2759, Sep. 2020.
- A. Largent, K. Kapse, S. D. Barnett, J. D. Asis-Cruz, M. Whitehead, J. Murnick, L. Zhao, N. Andersen, J. Quistorff, C. Lopez, and C. Limperopoulos, “Image quality assessment of fetal brain MRI using multi-instance deep learning methods,” Journal of Magnetic Resonance Imaging, vol. 54, no. 3, pp. 818–829, Sep. 2021.
- G. Ongie, A. Jalal, C. A. Metzler, R. G. Baraniuk, A. G. Dimakis, and R. Willett, “Deep learning techniques for inverse problems in imaging,” IEEE Journal on Selected Areas in Information Theory, vol. 1, no. 1, pp. 39–56, May 2020.
- W. Shi, G. Yan, Y. Li, H. Li, T. Liu, C. Sun, G. Wang, Y. Zhang, Y. Zou, and D. Wu, “Fetal brain age estimation and anomaly detection using attention-based deep ensembles with uncertainty,” NeuroImage, vol. 223, no. 117316, Dec. 2020.
- M. F. Beg, M. I. Miller, A. Trouve, and L. Younes, “Computing large´ deformation metric mappings via geodesic flows of diffeomorphisms,” International Journal of Computer Vision, vol. 61, no. 2, pp. 139–157, Feb. 2005.
- R. Heckel and P. Hand, “Deep decoder: Concise image representations from untrained non-convolutional networks,” arXiv:1810.03982v2, Feb. 2019, https://arxiv.org/abs/1810.03982.
- P. W. Holland and R. E. Welsch, “Robust regression using iteratively reweighted least-squares,” Communications in Statistics - Theory and Methods, vol. 6, no. 7, pp. 813–827, 1977.
- R. A. Maronna, R. D. Martin, V. J. Yohai, and M. Salibian-Barrera,´ Robust statistics, Theory and Methods (with R), 2nd ed., ser. Wiley Series on Probability & Statistics. John Wiley & Sons Ltd, 2019.
- C. S. Burrus, J. A. Barreto, and I. W. Selesnick, “Iterative reweighted least-squares design of FIR filters,” IEEE Transactions on Signal Processing, vol. 42, no. 11, pp. 2926–2936, Nov. 1994.
- D. C. Noll, F. E. Boada, and W. F. Eddy, “A spectral approach to analyzing slice selection in planar imaging: Optimization for throughplane interpolation,” Magnetic Resonance in Medicine, vol. 38, no. 1, pp. 151–160, jul 1997.
- A. Mang and L. Ruthotto, “A Lagrangian Gauss–Newton–Krylov solver for mass- and intensity-preserving diffeomorphic image registration,” SIAM Journal on Scientific Computing, vol. 39, no. 5, pp. B860–B885, Sep. 2017.
No competing interests reported.