(1)_format-1500w.png)
The Decades of Development in Computer-Aided Diagnosis (CAD): How it is Assisting in Mammography and Enganging in the Field of Radiomics
With the advent of artificial intelligence, “big data”, and machine learning, we are moving toward the rapid expansion of the use of these tools in the daily life of physicians, making each patient unique, as well as leading radiology toward the concept of a multidisciplinary approach and precision medicine. In this article, we will present the main aspects of the computational tools currently available for analysis of images and the principles of such analysis, together with the main terms and concepts involved, as well as examining the impact that the development of artificial intelligence has had on radiology and diagnostic imaging in mammography.
The genesis of computational decision-making in medicine started in the 1950s. The success with ‘automated computing diagnosis’, the initial idea of computers independently diagnosing conditions, was low in its early attempts. This was mainly due to the excessively high expectations from computers that were not sufficiently powerful or advanced in image-processing like the GPU’s (Graphics Processing Units) embedded in most radiological equipment today. Early studies on quantitative analysis disagreed with the established belief that computers may replace radiologists due to their high precision and efficiency in detecting abnormalities, thus the concept was deemed unachievable.
In 1960, R.L. Engle in his review article on 30 years’ experience stated in his conclusion “Thus, we do not see much promise in the development of computer programs to simulate the decision making of a physician. ---- However, after many years we have concluded that we should stop trying to make computers act like diagnosticians.”, Though it may seem like a ‘dead-end’ to the AI facilitation in the field of radiography, not all was lost. From this very unsuccessful scheme sprouted the excelling idea of ‘computer-aided diagnosis. Engle himself almost foreshadows this in his conclusion, and it is evident that in today’s world algorithm results are used as ‘second opinions’ to experts, and the machinery is not the expert itself. Fortunate as it is that ‘computer-aided diagnosis’ is more ethically approved and in balance to many values such as confidentiality and dignity that healthcare institutions uphold.
Systematic investigations that began in the 1960s, lead to the concept of utilization of ‘computer-aided diagnosis’ in a Picture Archive and Communication System (PACS) environment. This was initialised in the Kurt Rossmann Laboratories for Radiologic Image Research in the Department of Radiology at the University of Chicago in 1980. The basic strategy for the detection and quantitation of lesions in medical images is based on the same process radiologists’ use. Although seeming quite logical and straightforward, it limited the potential of the CAD system to be only as good as the radiologist, thus providing little assistance.
In Rossmann’s Laboratories where a basic scheme developed for the detection of microcalcifications in mammograms, the calibre of the program needed upgrading. The sensitivity of the scheme was 85% high, as expected of computational algorithms, but a significant complication arose as the numbers of false-positive results from the scan analysis were substantial. A quoted figure of 4 per mammogram, illustrated the uncertainty and unpredictability of the computerised schemes, making them unviable to be utilised by radiologists in clinical work. The scheme continued with an iterative improvement process to correctly identify the number of abnormalities and separate benign and malignant lesions. The journey to find the most effective and efficient ways to operate CAD, was indeed tumultuous, as you will acknowledge as you read on, but nevertheless, the scheme at Rossmann became a foundation of an international project between South Korea, UK & the USA, which will also be discussed in the later parts of this article.
CAD development rocketed internationally and began to supply the demand in many industries. As soon as the first CAD patent was filed in 1987, commercialised CAD products essentially became the ‘new kid on the block’ in the global medical market. These forms of CAD were simple and nothing as cognitively complex as CAD being developed in Rossmann’s laboratories or the CAD, we have now.
Representation of the function (matrix) of a grey-scale digital image (axial slice of a chest CT).
The fundamental of digital medical imaging is an f (x, y) function in a spatial coordinate-partitioned grey scale that can be represented by a matrix in which the intersection of each row and column identifies a single point (pixel) within an image. The value of each pixel in the matrix identifies the grey level at that point (x, y) on a scale of integer values that represent black (the lowest value), white (the highest value), and shades of grey (intermediate values).
As the CAD industry stabilised, after integrating its principles into common medical practice, an important question came into debate…” Can the development of CAD shift from ‘computer-aided’ to ‘automated computer diagnosis’?”. Ultimately, can we go back to the initial aim we had back in 1950 in the hope that it would be successful this time around? During two-panel discussion sessions in 2002 Paris and Montreal, about half of the participants voted for the possibility that CAD would be shifted to ‘automated computer diagnosis’ within 50 years, whereas the other half voted against this prediction. So far in time, we have expressed that we have the ability, but we shouldn’t let the possibility of it to occur, as it will complicate ethics and legality of how patients are being treated holistically and diagnosed.
With this being said it is crucial to clarify the basic procedure that CAD systems use in much radiographic equipment. Firstly, ‘Pre-processing’, where reduction of artefacts (bugs in images), and harmonization of image quality and contrast & filtering occurs to reduce the error in analysis. Following is the stage of ‘Segmentation’, where differentiation of structures in the image such as vessels and organs around the lesion are identified. Additionally, the matching of this image with those already existing in the anatomic databank to facilitate further analysis, and identify a ‘region of interest’ (ROI). When the image is analysed by the CAD system, it focuses on the ‘ROI’, more specifically the form, density, size, relative location, proportions, grayscale colour and neighbouring structures. For the last stage, the evaluation proceeds to classify and score the ROI using classifiers and neural networks. With the result and the analysis from the radiologist, the severity of the condition is classified using the
Image classification typically involves defining the image within a pre-established category, such as normal versus pathological. One of the most widely studied areas in artificial intelligence and image classification in machine learning. Machine learning allows the identification of patterns seen in previous cases and experiments from databases, as occurs with human intelligence. Machine learning methods have been applied to classifying images acquired with various imaging modalities.
However as advance and robust the software may be, CAD developers and users should be mindful of the importance of rigorous training, validation, and independent testing, as well as user training in clinical settings to ensure not only the generalizability of the standalone performance to the real world environment but also the effectiveness of clinicians using CAD in practice. With proper user training and understanding of the capability and limitations the deep learning technology, together with proper monitoring, objective assessments, and constructive feedback to enable further research and development, it can be expected that CAD technology will continue to progress and reach the goal of providing truly intelligent aids to improve health care.
Keeping this in mind, a retrospective study in January 2020, an AI algorithm was developed and validated with 170 230 mammography examinations collected from five institutions in South Korea, the USA, and the UK, including 36 468 cancer positive confirmed by biopsy. For the multicentre, observer-blinded, reader study, 320 mammograms were independently obtained from two institutions. 14 radiologists participated as readers and assessed each mammogram in terms of likelihood of malignancy (LOM), location of malignancy, and the necessity to recall the patient, first without and then with the assistance of the AI algorithm.
The AI Algorithm being tested here is using the largest breast cancer dataset for its training, thus resulting in improved diagnostic performances compared to radiologists, especially in early-stage breast carcinoma detection. Continuing from the fact that CAD systems have a large number of false positives with account for exhaustion and an increase in unnecessary additional examinations, this new CAD system seeks a new approach using deep learning in the form of descriptors.
Four-view heatmaps and an abnormality score per breast (ie, the maximum of the craniocaudal and mediolateral oblique abnormality scores) for each input mammogram
These descriptors are convolutional neural networks (CNN) architectures the one used here is the ResNet-34, which consist of 2 stages, stage-1 (low patch level training), and stage-2 (fine level tuning). From this, the algorithm provides a pixel-level abnormality score as a heat map, certainly rooted in the same principle as the function matrix but of a higher intricacy. The abnormality scores are floating-point values between 0 and 1.
Abnormalities can also be picked up on ultrasound using a modified algorithm The grayscale ultrasound image in a 57-year-old woman shown above with incidentally detected breast mass on screening examination shows an indistinct irregular heterogeneous hypoechoic mass (arrows) at the 9 o’clock position in the left breast. The analysis scoring on the BI-RADS lexicon was lower than the true value from general radiologists, yet breast specialists and the AI software diagnosed precisely.
The 14 radiologists graded mammograms using a probability of malignancy (POM), and a likelihood of malignancy (LOM) score, with one test using the AI – algorithm supporting the radiologists' conclusion, and the other test is based solely on the radiologists' analysis. The POM score measured from 0-100 with 0 least probable to be malignant and 100 the most. The LOM score was graded from 1-7 with 1 least likely to be malignant and 7 most likely. The POM score was used for evaluation and detection purposes and the LOM scores for evaluation of diagnostic performance. The LOM score was modified from the Breast Imaging Reporting and Data System (BI-RADS) categories, to be appropriate for receiver operating characteristic (ROC) analysis.
ROC analysis for AI-unaided and AI-aided diagnosis: Sensitivity and specificity of each individual (including AI standalone) are marked on each curve
As evident, the data with AI assistance becomes of higher precision, validity and reliability. Factors measured were sensitivity and specificity. With cases being confirmed by biopsy the more cancer diagnoses correctly done, the higher sensitivity/ specificity score between 0-1. Quoting data collected overall from all countries a 95% true positive rate. In Korea 97%, UK 94% &, the USA 95%.
However, throughout the study, it has been shown that calcifications in breasts lower specificity and dense parenchymal tissue lower sensitivity of the AI Algorithms analysis making the heterogeneous appearances of breast cancer harder to read. Apart from the difficulties in diagnosis due to these factors, it is true to say the software is robust, yet in terms of feasibility has to be proved through prospective clinical trials
Thanks to the high-quality multinational large-scale data, the AI algorithm consistently showed excellent performance in various validation datasets. Today it is commercially offered by iCAD and Hologic. It is likely that, in the future, some CAD schemes will be included together with other software for image processing in the workstations associated with some specific imaging modalities such as digital mammography, CT, and MRI.
As promising as the future of CAD seems, there are many challenges we must tackle first many relating to the computational logistics of the developing systems. There are algorithmic limitations as to how holistically the software can approach a condition it is analysing. Maximum input on a patient comes from electronic health records, additionally, confidentiality comes into question with how much data software should process. Most systems are generally designed to diagnose for a single condition, thus providing suboptimal results for patients with multiple, concurrent disorders.
Due to the massive availability of data and the need to analyse such data, big data is also one of the biggest challenges that CAD systems face today. The increasingly vast amount of patient data is a serious problem. Often the patient data are complex and can be semi-structured or unstructured data. It requires highly developed approaches to store, retrieve and analyse them in a reasonable time.
During the pre-processing stage, input data requires to be normalized. The normalization of input data includes noise reduction and filtering. Processing may contain a few sub-steps depending on applications. Basic three sub-steps on medical imaging are segmentation, feature extraction/selection and classification. These sub-steps require advanced techniques to analyse input data with less computational time. Although much effort has been devoted to creating innovative techniques for these procedures of CAD systems, there is still not the single best algorithm for each step. On-going studies in building innovative algorithms for all the aspects of CAD systems are essential.
There is also a lack of standardized assessment measures for CAD Systems. Moreover, while many positive developments of CAD systems have been proven, studies for validating their algorithms for clinical practice has hardly been confirmed. Other challenges are related to the problem for healthcare providers to adopt new CAD systems in clinical practice. Some negative studies may discourage the use of CAD. Besides, the lack of training of health professionals on the use of CAD sometimes brings the incorrect interpretation of the system outcomes.
The next steps for CAD would be emerging in the field of Radiomics. Radiomics has been described as an extension of CAD that associates the quantitative characteristics (features) of images with patient data and clinical outcomes, not only allowing the diagnosis to be made but also providing information regarding the prognosis and treatment response. Given recent advances in targeted treatment and immunotherapy, particularly in the treatment of malignancies, the need for a robust approach to imaging analysis has become clear, and radionics has the potential to provide this in a non-invasive, rapid, timely, and affordable manner. Radiomic analysis is a process of massive extraction of features from tens to hundreds of exams, inserting these features into databases with patient clinical information, allowing them to be shared and analysed.
The volume of health data has been growing at a rapid pace in recent years, characterizing what some authors call the “big data era” of health and those electronic data are available in large quantities in information systems of large hospitals and other health care centres.
More recently, maturation of computational models has provided support to the clinical decision-making and prognostic prediction processes. In this article, we have presented and discussed the main concepts related to computer-aided image analysis, including aspects of artificial intelligence applied to precision medicine.
Many believe that artificial intelligence, machine learning, computer-aided diagnosis and radiomics will change the way radiologists and other imaging specialists work and will likely, very soon, change the perspective that everyone in the health care field has on their work. However, some people fear that radiologists and other specialists will be completely replaced by computer algorithms. Although simple tasks and exams might be performed and interpreted entirely by such algorithms, the role of the physician in verifying/validating the outcome, making the clinical-epidemiological correlation, and determining the best treatment the regimen is unlikely to be threatened.
Artificial intelligence will certainly help “reduce the backlog” of exams; shorten the time to act in urgent cases; streamline interpretation and reporting; increase diagnostic confidence; make the image analysis more objective and reproducible; offer more reliable prognostic information; assist in the teaching and learning of imaging techniques; lead radiology definitively toward the concepts of precision medicine and multidisciplinary patient assessment.
References
Artificial intelligence, machine learning, computer-aided diagnosis, and radionics: advances in imaging towards precision medicine (12/21/2020)
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7007049/
A computer-aided diagnosis system using artificial intelligence for the diagnosis and characterization of breast masses on ultrasound (12/21/2020)
Computer-Aided Diagnosis in Medical Imaging: Historical Review, Current Status and Future Potential (12/21/2020)
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1955762/
Changes in cancer detection and false-positive recall in mammography using artificial intelligence: a retrospective, multi-reader study (12/21/2020)
https://core.ac.uk/download/pdf/304635795.pdf
Thumbnail Image Credit: National Cancer Institution