Approach-of-Artificial-Intelligence-and-Medical-Diagnosis

Healthcare has been an active area of research and development in this age of Artificial Intelligence (AI). Whether it is surgeons using robots to assist them in surgeries with more precision, accuracy and non-invasiveness; physicians using medical imaging and AI based technologies to diagnose diseases efficiently and accurately; or scientists using gene sequencing and gene-editing technology powered by AI to discover drugs and cures for even rare and fatal diseases. However, the most noticeable revolution seen in this golden age is the use of AI for medical diagnosis. Also, Artificial intelligence transforming the ways in which doctors diagnose diseases. Otherwise the approach of artificial intelligence in medical diagnosis plays an important role

What is AI in medical diagnosis?

AI in medical diagnosis is the use of algorithms emulating human recognition to analyze complex biological data. Also, AI algorithms must be able to perform reasoning without direct human involvement. For example, doctors educate through several years of medical schooling, where they taught about medical practices theory as well as through practical. They pass exams, receive grades, learn over the years and get over their mistakes. The way human educated and taught to be a doctor, similarly AI algorithms must learn to do their jobs.

AI algorithms can perform tasks:

Moreover, AI algorithms can perform tasks that require human intelligence such as image and speech recognition, also object detection and decision making, etc. However, AI algorithms learn what we want them to learn. Although We need to tell them explicitly what they should look for to complete the task at hand. We can train AI algorithms to automate complex tasks. Sometimes they can outperform humans in performing the jobs they are trained to do.

To develop an AI algorithm:

Further, To develop an AI algorithm, first data fed to the system along with appropriate labels or annotations that have to learned and recognized by the algorithm (Figure 1). The algorithm trained with enough dataset, also it can learn effectively. Once the training complete, accuracy of the algorithm evaluated to ensure its performance, similar to grades that are given to the students in their exams. Moreover, In order to evaluate the performance of the algorithm, unseen data or testing data, which has not been given during the training, is fed to the algorithm to assess its ability to determine the correct answer. If the results on unseen data are not up to the mark, the algorithm can be modified or trained with more data, to able to perform the designated task.

Figure 1 :
The working of an AI algorithm to segment the brain into different regions. The deep learning algorithm learns the anatomy of brain structures and regenerates the brain regions separated out from the MRI scan. The model learns the morphology of brain regions and can be used to segment brain regions from other brain scans. This application is useful for physicians in diagnosing brain disorders.

How it can be diagnosed?

Moreover, AI can assist physicians in every way, ranging from prognosis or diagnosis of disease and robot-assisted surgery to the discovery of personalized drugs and therapies. Imagine this! a fever for a few days and experiencing the symptoms of a cold. You have sore throat and your breathing is constricted. You go to the hospital. In the hospital, your doctor asks you about your symptoms and takes notes. Your doctor feeds these symptoms in a medical diagnosis system. Within a few seconds, this AI system retrieves your medical history, analyses your symptoms and presents your doctor with your recommended diagnosis. The doctor compares this AI diagnosis with his/her own diagnosis and discusses the results with you. This took much less time and improved the precision of your doctor’s diagnosis.

Cold is a common disease:

Although cold is a common disease and pretty much everyone knows about its symptoms. But think about other diseases such as neurodegenerative disease Parkinson. Here diagnosis is much more difficult and the patient’s survival depends on the efficiency and correctness of the diagnosis. An AI-based medical diagnosis, in this case, could do wonders for patients. This is the future of medical diagnosis where doctors can use AI-assisted technology to diagnose all kinds of diseases.

Algorithms in machine learning:

Moreover, There are a lot of algorithms in machine learning and specifically in deep learning that can learn from complex biological data. For medical diagnosis, several types of data are used by AI applications including (i) numerical or clinical data such as blood pressure and heart rate, (ii) medical images such as MRI, CT, X-rays, PET and biopsy tissue images and (iii) genomic data such as gene expressions, somatic mutations and copy number variations data. The type of data depends on the technology used for medical diagnostic tests.

Job of the AI algorithm:

Further, The job of the AI algorithm to learn from data to solve the tasks at hand such as identifying the stage of cancer,also identifying the presence of arterial clot, identifying the cancerous tissues, also classifying various stages of Alzheimer’s disease, detecting pneumonia in the lungs or segmenting the tumor affected regions of the brain. Nowadays, AI algorithms used to discover personalized drugs for diseases such as cancer where accurate diagnosis and therapy are crucial for the patient’s survival. In order to use an AI algorithm for medical diagnostic tasks, it must be compared with the performance of human experts such as doctors or physicians to determine its value and applicability in clinical practices.

Techniques and Theory :

In fact, AI has solved many clinical problems with the help of advancements in computation power and also huge medical data generated in clinical systems. Moreover, There are various recent applications of clinically accurate and relevant algorithms that can help in medical decision making. One such example is an existing system that outperforms clinicians in image classification tasks. Researchers at Seoul National University Hospital and College of Medicine[1] developed a system in 2018 using deep learning. Their system detects malignant pulmonary nodules and potential cancers by analyzing chest radiographs. The algorithms’s performance was compared with multiple physicians’ detections. In 17 out of 18 cases, their algorithm outperformed doctors in the evaluation of the same images


Figure 2: Medical scans of a 78-year-old female patient with a 1.9cm nodule at the upper left lobe. (a) In the chest x-ray, the nodule is slightly visible (indicated by arrowheads) and detected by 11 out of 18 physicians. (b) In the CT scans with contrast enhancement, lung adenocarcinoma is detected through biopsy (indicated by arrow). (c) The deep learning-based system was able to detect the nodule with a confidence level of 2, outperforming detection by 5 radiologists resulting in confidence elevation by 8 radiologists.
Image Source: Nam, Ju Gang, et al. “Development and validation of deep-learning-based automatic detection algorithm for malignant pulmonary nodules on chest radiographs.” Radiology 290.1 (2018): 218-228.

Google AI Healthcare:

In fact, Google AI Healthcare also contributed to this field by developing an AI-based medical diagnostic algorithm named as Lymph Node Assistant (LYNA)[2]. LYNA detects breast cancer nodal metastasis from lymph node biopsies by analyzing histology slides prepared by staining tissue samples. Although its not the first AI-based solution to analyze histology samples, LYNA is able to identify the cancerous regions unrecognizable by the human eye in the tissue samples. This system was tested on two datasets and obtained a receiver operating characteristic of 99% in classifying the tissue samples whether they are cancerous or not. Moreover, this system was used by doctors side by side with their traditional biopsy analysis. It was found that LYNA halved the time taken by doctors to review and analyze the tissue samples.


Figure 3: LYNA usage for hematoxylin-eosin–stained images. A slide can be analyzed in less than a minute on average on a cloud platform.
Image Source: Liu, Yun, et al. “Artificial Intelligence–Based Breast Cancer Nodal Metastasis Detection: Insights Into the Black Box for Pathologists.” Archives of pathology & laboratory medicine (2018).

Another fine example is an AI-based system:

Another fine example is an AI-based system developed recently by researchers at the German Cancer Research Center (DKFZ)[3]. With the use of deep learning algorithms, their system outperformed 136 out of 157 dermatologists in the classification of dermoscopic melanoma images. At the University of California in San Diego, Kang Zhang along with his team, developed an AI algorithm to diagnose roseola, chickenpox, influenza, glandular fever and hand, foot, mouth diseases with 90% to 97% accuracy. Their system was trained on 1.3 million patients’ medical records from a medical center in China.

AI can identify rare genetic syndromes:

In fact, Today AI can identify rare genetic syndromes by the features of someone’s face. People with genetic diseases sometimes have unique and different facial features. But there are hundreds of genetic conditions and diagnosis of these genetic diseases by associating them with facial features is complex and tricky. Moreover, At a biotechnology firm FDNA in Boston, Yaron Gurovich[4] and his team developed an AI system based on deep learning that can analyze facial images for overall impression or gestalt and outputs 10 genetic syndromes that a patient is most likely to have.

Their system was named DeepGestalt that was trained on 17000 images from patients with more than 200 types of genetic syndromes. DeepGestalt was tested on 502 images and obtained an accuracy of 91%. Moreover, the system was able to identify the genetic source or mutation associated with the physical appearance 64% of the time. Their system is helpful for doctors to understand the association between genetic makeup and facial phenotypes. 

Google AI system:

Moreover, Google AI system was developed for the diagnosis of lung cancer[5]. The system was trained on 42000 CT scans from patients and outperformed the diagnosis of 6 radiologists and elevated the detection rate of lung cancer by 5%. Another AI system IDx-DR[6] was approved by the FDA to detect diabetic retinopathy from the retinal scans. In the clinical study, the system detected diabetic retinopathy with 89.5% accuracy that leads to its approval. This tool is also able to perform detection without any interaction from the doctor.

Nvidia Clara:

Nvidia Clara[6] is playing the leading role of innovation in the field of medicine and healthcare. It is an AI-powered medical imaging platform that has greatly improved the time to process and analyze clinical images. Also, It provides timely and early diagnosis to physicians, specifically working in the ER departments of the hospitals. In fact, Most notable innovations are from Silicon Valley. In this regard, a startup, Subtle Medical[8] is developing tools to speed up and improve the productivity of radiologists by introducing AI-powered solutions. Imfusion[9] is another startup from Silicon Valley that is developing technology to construct 3D images from the 2D ultrasound.

Increasing the physicians’ ability to accurately and efficiently diagnose diseases:

Specifically, These examples indicate that AI based solutions are definitely increasing the physicians’ ability to accurately and efficiently diagnose diseases. On a short term basis, these algorithms can be used by doctors alongside their typical diagnosis to assist them in speeding up their work with increased confidence and accuracy. On a long-term basis, AI algorithms approved by the government can work independently in the clinics, enabling doctors to focus on the cases that cannot be solved by computers so far.

How it is equal to the human expert system

AI is equal to human expert systems in making medical diagnoses. Why and why not? With the advent of AI, it’s potential in the healthcare industry has caused excitement. Advocates are even saying that AI will decrease the strain on resources by freeing up the time for doctor-patient interaction. In this way doctor will be able to focus on much more complex cases that require human intelligence and have not been solved by machines so far. AI also has the potential to develop personalized drugs and treatments.

Seeing the revolution of IA in medicine and healthcare, the government of UK has announced £250 million for the establishment of a new AI laboratory in NHS.

Growing application:

 Moreover, The growing application is the use of AI algorithms in interpreting and understanding clinical images. In this aspect, the use of deep learning algorithms is common. Deep learning heavily relies on data and its quality in making diagnostic decisions. This technology has shown promising results in diagnosing several medical conditions ranging from cancer to genetic syndromes. A few examples of cases where AI outperformed human experts are:

  • System developed by Seoul National University Hospital and College of Medicine[10] to detect malignant pulmonary nodules.
  • Google AI Healthcare developed LYNA[12] system to detect breast cancer nodal metastasis from lymph node biopsies.
  • An AI-based system developed by researchers at the German Cancer Research Center (DKFZ)[13] for the classification of dermoscopic melanoma images.
  • An AI algorithm developed at the University of California in San Diego to diagnose roseola, chickenpox, influenza, glandular fever and hand, foot, mouth diseases.
  • DeepGestalt developed at FDNA in Boston[14] that can analyze facial images for overall impression or gestalt and outputs 10 genetic syndromes that a patient is most likely to have.
  • Google AI system was developed for the diagnosis of lung cancer[15].
  • An AI system IDx-DR[16] was approved by the FDA to detect diabetic retinopathy from the retinal scans.
  • Google DeepMind in collaboration with Moorfields Eye Hospital[17] has developed a system, that can identify 50 different eye diseases, matching eye specialists.

IA algorithms scale up to human intelligence:

Moreover, There are several examples where AI and deep learning have matched if not beaten, physicians and specialists in diagnosing diseases. But the question is how much these IA algorithms scale up to human intelligence. In order to find the answer to this question, researchers conducted a comprehensive study on the review of published work. The study was conducted at the University Hospitals Birmingham NHS Foundation Trust by Dr. Xiaoxuan Liu[18] and her team. The aim of the study was to compare the performance of deep learning algorithms versus healthcare professionals in diagnosing diseases from medical images. Researchers found that deep learning algorithms are on par with human experts in medical decision making.

Revolution made by AI in healthcare:

This study highlighted the revolution made by AI in healthcare. In this study, they focused on the research papers since 2012, a year when deep learning started to grow. They initially considered 20000 studies, only 14 of them were high quality research papers, where training was performed by using good quality data, used a different dataset for evaluation and showed the same images to human specialists. For these 14 studies, the accuracy of deep learning systems for disease detection was 87%, compared to 86% for human experts. However, the health professional was not given the additional patient information that affected their diagnosis. Dr. Xiaoxuan Liu, the lead author of the study said that although there are several AI systems claiming to outperform human experts, yet these systems can at best be equivalent.

Pros and cons:

In fact, By seeing the advancements made by humans in the field of medicine and health care, we can anticipate that the future of AI-powered medicine is not far, where a patient can see a computer before going to the doctor and doctors can rely on AI diagnostic systems for making decisions. AI is revolutionizing traditional medical practices. The advancements in AI have made it possible for the days of misdiagnosis to move behind us by not only empowering diagnostic practices but also the treatment of diseases from their root causes rather than symptoms. The days of traditional blood pressure measurements gone. There were days when you would need a large storage space to fit a 3D scan of an organ on your laptop, and thinking about processing it would be a nightmare.

Large data storage:

Today, however, large data storage, as well as its processing, is not an issue. The data generated in clinical tests stored in electronic medical health records, where physicians can keep track of your medical history. Advanced medical imaging technologies coupled with AI techniques allow for high-performance data-driven medical applications. Although Development of next generation sequencing technologies is revolutionizing the discovery and development of drugs. Hence, AI applications have drastically transformed and will continue to transform the ways of clinical problem-solving.

What is holding them back from clinical and practical use?

      There are several examples of works that highlight the potential strengths of AI algorithms in medical diagnosis. Although various AI algorithms compete or sometimes outperform human experts, they have not integrated into medical practices yet. Why? Because, although AI technology has impacted medical interventions in a meaningful way, however, there are various regulatory implications and concerns that need to be addressed first.

AI algorithms is a challenging task:

            Regulating AI algorithms is a challenging task. The US Food and Drug Administration (FDA) has signaled approval of some medical devices[19], however, approval guidelines do not exist currently. There is another concern, the people developing medical applications usually not clinicians, thus they might need to learn a lot about medicine to develop a fully functional AI system. On the other hand, clinicians using AI applications practically might need to learn using the AI system effectively as well as its limitations. There is a need for FDA approved standards and guidelines that can be used to specify the requirements for algorithms and their clinical deployment.  

AI is able to help in diagnostic and basic clinical tasks:

            Today, AI is able to help in diagnostic and basic clinical tasks, however, fully automated brain or heart surgeries are not possible yet. Similarly, AI is being used in medicine[2] but its applicability for patient care is limited. AI medical systems also face difficulties to approved and trusted by patients. Without a clear understanding of AI technology and its capabilities, patients might reluctant or not willing to let it used to fulfill their medical needs. Would an AI algorithm outperforming human experts might approved by patients to be used in clinical practices? This is an open question to answered and it impacts the confidence over the decision-making power of AI systems.

Ability of AI algorithms:

 In fact, The ability of AI algorithms depends on the data that fed to them. Correct data is vital for the correct functioning of the AI system. Also Its possible that developers of AI systems might not know about the data and they feed incorrect or misleading data to the algorithm, causing medical malpractice. This can avoided by reducing the communication gap between clinicians and developers. If clinicians inform programmers about the data and its correct usage, errors can be avoided. Although We can say that the programmers need to properly understand the data and clinicians need to properly understand the limitations of the algorithms. This mutual understanding can help to build up confidence in AI systems and their applicability in clinics. These are the challenges need to addressed to increase the precision and efficiency of medical practices.

Blog Description :

In fact, AI has been an active part of the healthcare industry in recent years. AI can assist physicians in every way, or ranging from prognosis or diagnosis of disease and robot-assisted surgery to the discovery of personalized drugs and therapies. The most noticeable revolution seen in this golden age is the use of AI for medical diagnosis. A lot of algorithms in deep learning that can learn from complex biological data. Although, The job of the AI algorithm to learn from data to solve the tasks at hand. In fact, In order to use an AI algorithm for medical diagnostic tasks, it must compared with the performance of human experts to determine its value and applicability in clinical practices.

Although various applications of clinically accurate and relevant algorithms that can help in medical decision making. In fact many several AI systems claiming to outperform human experts, yet these systems can at best be equivalent and they have not integrated into medical practices. There are various regulatory implications and concerns that need to be addressed first such as regulating AI algorithms, or the need for FDA approved standards, patient’s trust, dependability on data and communication gap between the clinicians and developers. These challenges need to addressed to increase the precision and efficiency of medical practices.

Although there are several AI systems claiming to outperform human experts, yet these systems can at best be equivalent and they have not integrated into medical practices. Many various regulatory implications and concerns that need to be addressed first such as regulating AI algorithms, the need for FDA approved standards, patient’s trust, dependability on data and communication gap between the clinicians and developers. These challenges need to addressed to increase the precision and efficiency of medical practices.


Refrences:

[1] https://news.bloomberglaw.com/tech-and-telecom-law/fda-signals-fast-track-approval-for-ai-based-medical-devices-1

[2] https://www.nature.com/articles/s41591-018-0300-7


[3] Nam, Ju Gang, et al. “Development and validation of deep learning–based automatic detection algorithm for malignant pulmonary nodules on chest radiographs.” Radiology 290.1 (2018): 218-228.

[4] Liu, Yun, et al. “Artificial Intelligence–Based Breast Cancer Nodal Metastasis Detection: Insights Into the Black Box for Pathologists.” Archives of pathology & laboratory medicine (2018).

LEAVE A REPLY

Please enter your comment!
Please enter your name here

8 + 9 =