Fundamentals of artificial intelligence
Artificial Intelligence (AI) requires human intelligence and is performed by a computer.
An algorithmic model is developed by humans (programmers)—a code instructing the computer
step by step for necessary acting, reasoning, and learning.
Machine learning (ML) is an algorithm that can learn from data and act according to
this knowledge without extensive prior programming. Here, a specific task can be performed
based on particular patterns and inferences. It does not need any clear instructions
or any programs. Based on “training data,” a mathematical model is created and the
computer can learn to make predictions. The ML model performs better with exposure
to more data over time. The three main types of ML algorithms are:
-
(i) Unsupervised ML: It depends on finding patterns. Unlabeled data are given as an
input, following which the system looks for patterns and classifies the data according
to the identified patterns.
-
(ii) Supervised ML: It predicts algorithms based on past learning. Labeled data are
already given as input; hence, the system categorizes various new given inputs based
on what has been learned with old labeled data. One of the most commonly used forms
of supervised learning is deep learning.
-
(iii) Reinforcement ML: It uses a system of reward and punishment for algorithmic
training.
Deep learning (DL) is a subdivision of machine learning methods created on artificial
neural networks. This model is built based on how a huge data/information is processed
by the human brain. A well-designed and trained DL model can predict and perform classification
tasks with more accuracy, which exceeds human expertise at times. In the field of
medicine, AI can be divided into two branches: virtual and physical.[1]
Virtual branch of AI: It is primarily based on ML, where mathematical algorithms help
improve experience-based learning. The applications of virtual AI in medicine include
electronic health record (EHR) systems and neural network-based guidance in treatment
decisions.
Physical branch of AI: It is the use of machines (robots) in physical form (macroscopic
or microscopic) for assistance in performing daily tasks, surgeries, or delivery of
chemotherapeutic drugs to a specific site.
-
(1) Carebots: These are advanced and sophisticated robots. They are used to assist surgeons
during operation, as intelligent prostheses for handicapped people, and care of the
elderly.
-
(2) Nanobots: They refer to miniature devices (size = 0.1–10 µm) in the form of switches,
motors, shuttles, or cars carrying a particular drug or molecule. One such example
is the targeted drug delivery using nanoliposomes by overcoming the permeation and
diffusion barriers of the conventional therapeutic agents.
Applications of AI in oncology
AI has been and is being explored in various fields of oncology related to histopathology,
imaging, radiation oncology, medical oncology, translational oncology, and clinical
decision-making as depicted in [Fig. 1].
Fig. 1 Applications of artificial intelligence (AI) in oncology.
Cancer Screening
AI has found its vital place in the screening of cancers. Several convolutional neural
network (CNN) models from the 2017 Kaggle Data Science Bowl demonstrated 80 to 95%
accuracy for lung cancer screening by evaluation of suspicious nodules in thoracic
computed tomography (CT) films.[2] In the same year, the digital mammography (DM) DREAM challenge by IBM, tested whether
AI algorithms could be equal to or supplement the radiologists in interpreting mammograms.
AI-based algorithms have reviewed 640,000 digital mammograms with 81% and 80% specificity
and sensitivity, respectively.[3] Scott et al. assessed the performance of AI in predicting breast cancer from a large
dataset of mammograms from both the USA and UK.[4] The false-positive and false-negative rates reduced by 5.7%, 1.2% (USA, UK) and
9.4%, 2.7%, respectively. An independent study was done comparing AI algorithms and
the radiologist's interpretation of screening mammograms for detecting breast cancer.
The AI surpassed all human readers (six radiologists): the area under the receiver
operating characteristic curve (AUROC) for the AI system was greater by an absolute
margin of 11.5% than that for an average radiologist.
Cancer Diagnosis
Within the field of diagnostics, AI is gaining a pivotal role in accurate and speedy
results.
Clinical Images
The CNN-trained algorithm using 130,000 skin images classified malignant lesions with
higher sensitivity and specificity as compared with 21 dermatologists.[5] A 94% sensitivity in polyp detection using colonoscopy images from 1,290 patients
was observed with CNN.[6]
Radiographic Imaging
A study using a DL-based AI algorithm for the prediction of disease based on electronic
health records (EHRs) of patient's clinical history, laboratory investigations, and
imaging findings has been successful in this arena. In this, a novel framework called
“deep patient” was developed based on a large dataset of patients' EHRs. Different
patient sets with common patterns in the given data were identified. The new data
were entered later and tested for its accuracy in predicting novel diseases in the
next 1 year. It predicted the development of a variety of diseases including cancers
of the prostate, rectum, and liver with 93% accuracy.[7] Another CNN-based model showed >85% accuracy in identifying extranodal extension
(ENE) in head and neck cancers on diagnostic, contrast-enhanced CT scans.[8]
The CNNs have been able to predict both IDH mutation and MGMT methylation status with 85% to 95% and 83% accuracy, respectively, based on raw imaging
data alone.[9]
[10] A CNN model successfully predicted the complete response to neoadjuvant chemoradiation
with 80% accuracy in locally advanced rectal cancer.[11] Additionally, a radiomics signature using extracted features from CT data and an
ML algorithm were able to predict underlying CD8 cell tumor infiltration and, remarkably,
response to immunotherapy for a variety of advanced cancer patients in phase 1 trials
treated with anti-PD-1 or anti-PD-L1 monotherapy.[12]
Digital Pathology
The diagnostic rates of lymph node metastasis in resected specimens of carcinoma breast
using DL CNN algorithms were found to be the same as those observed in a group of
pathologists.[13] Similarly, in adenocarcinoma of the prostate, a 75% rate of agreement of Gleason
grading was seen between the DL algorithm and pathologists.[14] A trained CNN was developed to predict six genetic mutations (STK11, EGFR, FAT1, SETBP1, KRAS, and TP53) from lung cancer biopsies. This paved the way for predicting a genotype or mutation
based on histopathologic architectural patterns.[15] These methods will be cost-effective as compared with tests for direct mutational
analysis.
Radiation Therapy
AI has been proposed for image acquisition, tumor segmentation, image registration,
radiation planning, and radiation delivery—the steps during treatment workflow in
radiotherapy.
Image Acquisition
CT scan is the basic imaging for getting electronic density values. It helps in the
planning of radiation therapy using algorithms for the calculation of doses. However,
MRI is advantageous for better soft tissue delineation and acquisition of multi-planar
images. There has been continuous research in the development of methods to generate
a CT scan using MRI data—a synthetic CT scan (sCT). Of these methods, deep embedding
CNN has shown promising results in terms of efficiency, time consumption, high-quality
image resolution, and fewer artifacts.[16] Algorithms for creating sCT from primary MRI images and for generation of radiation
plan have been used in MRI only for prostate radiotherapy.[17]
Segmentation of Tumors and Organs at Risk
Manual contouring of organs at risk (OAR) and the volume of the target is a time-consuming
task with huge interobserver variability.[18] To overcome this hurdle, automatic contouring software using knowledge-based algorithms
(atlas, machine learning, models based on statistical shape and appearance); region-based
(adaptive thresholding, graph cuts, and contouring by watershed); or a combination
of the knowledge and region-based have been used.[19] In a recent study by Lustberg,[20] the deep learning contouring outperformed the atlas-based contouring for the lungs
and spinal cord. With deep learning methods, 79% of the median time was saved in comparison
to manual methods. Men et al.[21] hypothesized a new deep dilated CNN-based method for target auto-segmentation and
volume delineation of OAR. It took on an average 45 seconds for one patient for segmentation
of all clinical target volume (CTV) and OAR which is far less time than that taken
for drawing the structures conventionally.
Image Registration
For image registration (to align an image to the reference image), mathematically
applied transformations are used. The two traditional registration methods used in
radiotherapy are the intensity-based method and the rigid method. In their review
on newer techniques of image registration, Viergever et al[22] found the deep learning process for registration of the images to be easier and
user-friendly. This has been substantiated by studies done later by Yang et al.[23] and Miao et al.[24] using DL- and CNN-based methods, respectively.
Radiation Planning
The planning process for radiotherapy using AI algorithms was pioneered by McIntosh
et al.[25] They used a voxel-based method for predicting the dose and dose-mimicking method
in planning radiation for head and neck cancers. The advantage of this adaptive radiotherapy
planning process is not limited to saving time but also includes dose adjustments
related to age, sex, race, and genetic makeup.[26]
PORTOS is the first clinical radiogenomics assay that uses predictive biomarkers to
determine the tumor sensitivity to radiation.[27]
Radiation Delivery
The recent advances in radiotherapy such as IGRT and IMRT focus more on image-guided
delivery of radiation and less on the patient's position. Positional deviation leads
to dose variation and hence compromises the efficacy of the treatment. Ogunmolu et
al.[28]
[29] developed a soft-robot actuator for head and neck radiotherapy without using a mask
to monitor intra- and interfraction movements during radiation. It showed promising
results in controlling the movement of the head to within the nearest 2 mm (millimeter)
as compared with the trajectory of reference. Park et al.[30] have used the data of breathing patterns for intra- and interfraction fuzzy deep
learning (FDL) to decrease tumor-tracking time and adjust radiation doses in lung
cancer patients according to breathing movements.
Treatment Outcomes
In 2003, an ANN analysis was done on 125 nonsmall cell lung cancer patients to predict
their 5-year survival rates based on the immunohistochemical and clinicopathological
variables.[31] The ANN-based prediction model was superior to logistic regression analysis in estimating
the 5-year NSCLC survival rates (87% vs. 78%). In-silico analysis using clinical and
dosimetric data was also done to predict the genitourinary and lower gastrointestinal
toxicity resulting from radiotherapy to prostate,[32]
[33]
[34] hepatobiliary toxicity after liver irradiation,[35] and rectal toxicity in cervical cancer patients receiving radiation therapy.[36]
Translational Oncology
The utility of DL neural networks has extended even to the field of translational
oncology. With advancements in biotechnology and biochemistry, there is huge information
on protein structure, and protein–protein and drug-protein interactions. The structure
of a protein can be predicted.[37] To speed up the drug development process, CNN and ML methods were used to predict
the side effects of various drug combinations.[38] It saves time and financial resources by halting the drug development process due
to serious toxicities. The cells can be characterized according to different mitotic
stages [39] based on microscopy images. The progenitor cell lineage in the future can be delineated
based on their microscopic images.[40] Deep learning artificial neural networks (DL ANNs) have been trained to predict
the chances of failure of around 200 drugs based on transcriptomic response signatures.[41] By combining both genomic and chemical characteristics, AI can predict the sensitivity
of tumor cells to therapeutics.[42] The role of AI has been extended even to immunologicals by predicting the peptide
major histocompatibility complex (MHC) binding of the immunological drugs using CNNs.[43]
Role of AI in Clinical Decision-Making
With the enormous amount of ongoing research in oncology that includes clinical trials,
new drug development, and the discovery of predictive biomarkers, there is a splendid
opportunity for AI to assist in data synthesis and guide us in making and taking decisions
for cancer patients. More commercial applications are being developed that utilize
DL and natural language processing to utilize this huge data from bench-side to bed-side.[44] International business machines (IBM) in collaboration with Watson for Oncology
(WFO) used AI for linking the patient data to national treatment guidelines. The WFO
demonstrated high concordance rates of the multi-disciplinary team recommendations
with that of AI in breast cancer patients.[45]
[46] However, the same IBM Watson cognitive computing system failed in taking the right
decisions related to cancer treatment. After a huge expenditure of $62 million, the
joint venture of IBM and University of Texas MD Anderson Cancer Center for the development
of an advanced Oncology Expert Advisor had to be called off! Retrospection showed
training with a small dataset as a cause for this failure.
Role of AI in Hemato-Oncology
Several CNN-based models have distinguished between different white blood cells on
peripheral smear with more than 95% AUROC scores.[47] Similar work is underway for the interpretation of bone marrow specimens. In patients
with lymphoma, automated analysis for bone marrow involvement has been successful.[48] The risk stratification for advanced Hodgkin lymphoma based on PET-CT images was
also tried with encouraging results.[49]
In malignant hematology, there are myriad clinical and laboratory-based prognostic
scores for each cancer. However, there is a significant variation on a case-to-case
basis. AI can deal with this complex data and help in personalized therapy. Upfront
AML/MDS risk stratification can be improvised with the help of AI.[50] One of the most important poor prognostic factors after treatment completion in
hematological malignancies is the detectable minimal residual disease (MRD). AI matched
the human-level performance in MRD detection using flow cytometry in patients with
acute myeloid leukemia.[51] A retrospective data analysis by Sasaki et al in CML-treated patients showed a longer
survival rate for ML-informed treatments. Algorithm-based models in MDS predicted
the response to hypomethylating agents.[52]
Role of AI in Resource-Limited Settings
In developing countries such as India, there is a huge disparity between the numbers
of patients and the numbers of health workers. A SWOT (Strengths, Weakness, Opportunities,
Threat) analysis for this cutting-edge technology in developing countries is as follows:
-
Strengths: Availability of large amounts of data, emerging young population with a
lot of talent and interest in this new technology, eco-friendly.
-
Weakness: High cost, lack of trained personnel, lack of standardization of data collection,
storage and processing, data privacy, data security, digital divide (lack of access
to high-speed broadband networks in rural areas).
-
Opportunities: Public–private partnerships, national and international collaborations,
Government funding (NITI–National Institution of Transforming India, DBT–Department
of Biotechnology) for the development of AI in health care settings.
-
Threats: Loss of technological and health care workers' employment, legal and ethical
issues.
Limitations of AI
This groundbreaking technology has its limitations that are hindering AI to be used
in routine clinical settings.
First, the greatest hindrance is the external validation of DL applications and the
generalizability of DL applications to all patients and tumors. The heterogeneous
medical data across institutions and different patient populations will require multiple
validation sets to prove the performance of AI. Second, access to data is a problem.
It would be difficult to provide data for diseases with less prevalence. Contributing
to this data scarcity, there are ethical and legal issues involved in sharing of protected
information about patient's health across different institutions and also data heterogeneity
along with incomplete data collection and competition between institutions. The use
of findable, accessible, interoperable, reusable (FAIR) data,[53] and many other opportunities are being provided for various research groups to address
this issue.[54] Third, tracing the exact logic behind the predictions of DL is a hilarious task.
Hence, it is rightly called the “black box” problem, which means the explanation for
how and why it gave the output based on the input given, cannot be provided by AI.[55] The era of personalized and evidence-based medicine demands a rationale for every
clinical decision being taken. The specific characteristic(s) of the input data that
contributed to the ultimate output cannot be pinpointed in DL algorithms. From the
perspective of both practitioner and regulatory bodies, this interpretability is a
challenge for adopting AI-based algorithms into the healthcare system.[56]
[57] Fourth, AI in radiotherapy and pathology is mostly dependent on the images for auto-segmentation,
hence quality and the number of images used for training dataset matters a lot in
giving the final output. Poor-quality images and fewer images lead to faulty assessments
by AI. Fifth, medical professionals lack the education and expertise required for
dealing with AI-related algorithms. A collaboration between departments of clinical
oncology, bioinformatics, and data science is needed to tide over this problem. Lastly,
being in an initial start-up stage, the cost will be higher and institutions need
financial planning accordingly.
Conclusion
The breakthrough technology of this millennium “artificial intelligence” has a promising
future in all the fields of oncology. It is a necessary tool for handling enormous
clinical data, accurate delivery of treatment, improving personalized treatment selection,
and predicting patient and disease outcomes.
Like any new invention, AI also has had and continues to have some hiccups, but more
research into these areas can help address them. It is not surprising to say that
in near future, AI will become a new “sixth sense” for every oncologist.
