• Raveena Rajpal

Deep Learning in Medical Image Analysis


Introduction

With the speedy development of computing (AI) technology, the utilization of AI technology to extract clinical knowledge has become a significant trend within the medical business. Victimization advanced AI algorithms for medical image analysis, one of the foremost vital elements of clinical designation and decision-making; it's become a vigorous analysis field in business and studies. The newest applications for in-depth learning in medical image analysis embrace a range of computer-related functions like designation, detection, classification, and registration. Among them, segregation, acquisition, and segregation square measure the foremost basic and widely used procedures. Medical image analysis is an efficient field of machine learning analysis, partly due to the relatively designed and labelled information. It's probably that this may be the place where patients initially have interaction with active intelligence systems. This can be vital for two reasons. First, supported actual patient metrics, medical imaging analysis is also a litmus (test | acid-base indicator) test of whether AI systems can truly improve patient outcomes and survival. Second, it provides a bed to check human interaction with AI; however, receptive patients will be in life-changing decisions, or assisted by a non-human character.


History of Medical Image Analysis

The symbolic AI paradigm of the 1970s led to the development of legal principles, professional programs. The first medical initiative was Shortliffe's MYCIN program, which promoted various antiretroviral therapies for patients. In line with this development, AI algorithms range from heuristics-based techniques to hand-drawn features to supervised learning strategies. Unconverted machine learning methods are also being researched, but most of the published algorithms use supervised learning methods, namely Convolutional Neural Networks (CNN). The widespread use of CNN in image capture occurred after Krizhevsky et al. won the 2012 ImagenetLarge Scale Visual Recognition Challenge (ILSVRC) by CNN which had a 15% error rate. There has been a dramatic growth in CNN architecture and applications research; now, CNNs have become the leading architecture for medical imaging.

Neural networks

  • Convolutional Neural Network (CNN) It is a widely used architecture in Medical Imaging. CNN may be a popular machine learning algorithm for visual comprehension and visual learning activities, because of its unique feature of maintaining visual image relationships. This captures the connection of the key element within the image and reduces the number of parameters that the algorithm has got to calculate, which increases computer efficiency. CNN can capture input and process both 2-dimensional images and 3-dimensional images with minimal changes. This is often a useful advantage in designing a hospital-based system, as some methods like X-ray have two sides while others like, CT or MRI scans are 3-dimensional. CNN and Recurrent Neural Networks (RNNs) are samples of supervised machine learning algorithms, which require significant amounts of coaching data. Unchecked learning algorithms are also explored in medical image analysis, such as Autoencoders, Boltzmann Limited Borders (RBM), etc.

  • Medical Image Segmentation Medical image segmentation is the technique of detecting organs or lesions in CT scans or MRI pictures, which can provide important information about the organs' shapes and volumes. This approach employs Convolutional Neural Networks (CNNs). CNNs operate similarly to a standard feedforward neural network, but they are considerably better able to cope with images since they incorporate a combination of techniques such as convolutions, max-pooling, and so on. The basic idea behind employing CNNs is to take a 2D input image and apply 2D filters to it (through a 2D CNN). Another method is to employ Transfer Learning, which involves training models with pre-trained state-of-the-art models and freezing the last few layers for learning weights appropriate to the situation.

  • Image Detection Deep Learning helps in the detection of cancer and diabetic retinopathy. Diabetes mellitus may be a disorder during which the pancreas fails to supply proper insulin or the body's system doesn't respond well to insulin, resulting in elevated blood glucose. Because of the shortage of experts in this field, manual detection of diabetic retinopathy may be a complex and time-consuming process. Automatic DR detection based on in-depth learning models has proven its best accuracy. Deep convolutional neural networks are widely accustomed to detect DR.

Application Areas

  • Cardiac - CT and MRI are commonly used medical imaging techniques to diagnose abnormalities of function and structure of the circulatory system. Automatic analysis of images from the above methods can help doctors study. The structures and performance of the guts muscle determine the explanation for a patient's coronary failure and see potential tissue damage.

  • Intracranial aneurysm - Intracranial aneurysm may be a common life-threatening disease usually caused by trauma, a vascular disease. Fracture of the aneurysm may be a critical event with high rates of death and morbidity. Therefore, accurate diagnosis of intracranial aneurysms is additionally important. Shi et al. has proposed an in-depth 3D modelling study to detect aneurysm in CTA images.

  • Brain Tumor - The in-depth study method accurately identifies and separates the brain tumour region from each brain MRI image. The magnitude and site of the ROI established a correct diagnosis of brain cancer.


Notable Companies Employing Deep Learning and AI in Medical Image Analysis (Healthcare)

Arterys, Inc.

Arterys is one of the most advanced artificial intelligence (AI) technologies for medical imaging. This idea makes use of cloud computing to perform some of the quickest medical calculations conceivable. The medical imaging analytics platform is well-known for its exceptional imaging speed and quality.


VoxelCloud

VoxelCloud uses artificial intelligence (AI) and cloud computing to provide automated medical picture analysis and diagnosis support. VoxelCloud develops computer-based discovery frameworks for recognising and diagnosing a variety of pathologies, including cardiovascular and pulmonary diseases, as well as eye disorders. Its solutions rely on cutting-edge computer vision, deep learning, and artificial intelligence technology.


Butterfly Network, Inc.

Butterfly Network is a digital health group with the goal of democratising healthcare by making medical imaging more widely available and affordable. 4.7 billion people throughout the world currently lack access to this basic technology. A cross-disciplinary combination of semiconductor engineering, AI, and connected mobile software drives its complete stack approach to deal with the problem. Butterfly iQ, the world's first handheld complete body ultrasound framework, is the result.


Infervision

Beijing based Infervision is a high-tech artificial intelligence company that uses deep learning to assist doctors diagnose medical images more effectively and precisely. Infervision successfully integrates many types of medical data into therapeutically valuable solutions and enhances precision analysis in the medical area, particularly in assisted picture diagnosis.


Zebra Medical Vision

The Imaging Analytics Platform from Zebra Medical Vision enables medical services businesses to identify individuals at risk of disease and provide preventive treatment routes to improve patient care. The company's AI solutions evaluate a large amount of clinical imaging data in real time, recognising various medical indications, allowing it to be the only AI Medical Imaging company with such a broad product offering.


Challenges and Future of Deep Learning in Medical Image Analysis

Although in-depth learning models have achieved great success in medical imaging analysis, small medical data sets are still a serious challenge during this field.

Inspired by the concept of transfer, another possible way is to form a website transfer that's familiar with a model trained in natural images in medical image applications or from one image mode to a different. Another possible option is integrating learning, where training is often done collaboratively between multiple data centres. Class imbalance is another major problem for medical image analysis. Numerous studies on the planning of the novel loss function, like focus loss, variable loss, and triple loss, are proposed to deal with this issue.

New research areas include forecasting, content-based image retrieval report or caption production, as well as material manipulation of LSTMs and reinforcement learning involving surgical robots.


Conclusion

The rise of advanced learning methods has made great strides in analyzing the medical picture with high accuracy, efficiency, stability, and balance. In this article, we've reviewed the newest developments in CNN-based in-depth learning strategies in clinical programs that include image classification, object acquisition, classification, and enrollment. An in-depth analysis of the pictures supported the diagnostic app on the physical body's four major systems, including the systema nervosum, circulatory system, gastrointestinal system, and skeleton, was reviewed. To be more specific, state-of-the-art is discussed during a sort of disease including brain diseases, heart condition, and osteoporosis.


References

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6945006/ https://www.analyticsinsight.net/top-10-medical-imaging-companies/ https://medicalfuturist.com/top-artificial-intelligence-companies-in-healthcare/ https://pubs.rsna.org/doi/full/10.1148/rg.2017160130

https://www.hindawi.com/journals/cin/2021/6215281/

https://spj.sciencemag.org/journals/hds/2021/8786793/

https://ieeexplore.ieee.org/document/8241753

tm.amegroups.com/article/view/36944/html

https://medium.com/the-research-nest/deep-learning-models-for-medical-image-analysis-and-processing-a4f8ba58e58f