Treatment in oncology has progressed dramatically over the last two decades with the advent of personalised biomarker specific targeted therapies typically benefiting metastatic treatment settings. However, a challenge all clinicians still face is the asymptomatic nature of common cancers such as breast and lung in the early stages of the disease. The ability to diagnose and treat potentially malignant tumours in their infancy irrefutably improves a patient’s potential clinical outcomes. With improvements in artificial intelligence (AI) software accurate and consistent diagnosis based on medical imaging may, in the near future play a more substantial role within public healthcare.
Historically imaging technicians (radiologists) have used medical imaging to identify, diagnose and survey tumour progression somewhat subjective, making it difficult to apply consistent statistical analysis based on these inputs to accurately diagnose patients or predict a prognosis. AI software such as deep learning (using a deep neural network) consistently indiscriminately recognises patterns in patient scans to generate a purely quantitative analysis across all individuals without human acquired bias sullying the results. This enables comparative analysis of past patient prognosis in relation to their history of medical imaging, which can be leveraged as a basis for predicting the prognosis for new patients based on their past and current scans. This provides a high degree of reliability and accuracy.
A deep neural network is a type of machine learning methodology. Using historic “training” data the algorithm builds a decision network, with probabilistic weightings based on the “seen” data functioning similarly to the neural pathways of the brain and hence the name. The network model synthesised by the learning algorithm, is tested with historic data which the algorithm has yet to encounter, in order to validate the outputs of the model. Provided that the model has performed well in the test phase, it can now be used predictively. Scans with unknown outputs are fed into the model which, based on its calculated expected outcome can be used to drive a decision.
Studies into the implementation of deep learning have progressed rapidly over the last decade. One such paper analysing the use of AI systems for breast cancer screenings has shown promising results for the future of AI utilisation in medical imaging screening . Identification of breast cancer in the earlier stages of the disease increase the likelihood of successful treatment. The study carried out by Mackinney, S. M. et al aimed to diagnose breast cancer by analysing mammogram images obtained from significantly sized UK (25,856 patients, 2012-15) and US (3,097 patients, 2001-18) databases. A statistically significant difference in the instance of false positives (reduction of 5.7% (USA) and 1.2% (UK)) and false negatives (reductions of 9.4% (USA) and 2.7% (UK)) was identified by the AI software in comparison with the participating radiologist’s initial decisions . In UK a system involving a second radiologist (reader) is used to verify the findings of the initial reader. It was found that the AI system did not outperform this error checking benchmark; however, it is estimated that the use of the AI system could reduce the second readers workload by approximately 88% (based on a scenario where the initial reader checks images concurrently with the AI software and utilising a second reader to check images where the AI and initial reader have not agreed) . This study represents a bright future for the use of AI in the diagnosis and identification of breast cancer and the potential for its uses amongst other tumour types.
The volume of medical imaging use across the UK and US is growing at an unprecedented rate. The volume of imaging is almost out of line with the budget growth of both US government and third-party payers costing structures. This has the potential to result in suboptimal healthcare for patients . In the US third-party payers have already begun implementing measures to cut imaging costs be discouraging hospitals from performing CT scans using pay-for-performance contract to incentivise hospitals to reduce imaging utilisation in exchange for an at-risk bonus . Furthermore, Medicare and Medicaid have, as part of the Deficit Reduction Act put into place a 25% reduction on payments for outpatient CT scans for any additional body part examined within one session in attempt to discourage use of imaging and lower payer costs; However CT examinations routinely involve more than one body part .
Aside from cost, the continuously increasing workload also puts strain on the limited number of radiologists tasked with assessing medical images. One study based on a single healthcare centre in the US estimated that the average radiologist has to evaluate an image every 3-4 seconds in an 8-hour shift to meet the workload . The efficiency of using intelligent software overseen by professional radiologists has the potential to revolutionise the determination of diagnosis and prognosis in oncology, by decreasing costing, workload and improving accuracy of medical imaging assessment.
AI in oncology is in its adolescence, with deep learning soon to reach maturity in the coming decade. The potential for improvements in clinical outcomes, reduction in treatment required for patients due to early diagnosis, costs saved through increased treatment and diagnostic efficiencies are but a few of the prospects that AI based solutions may enable. Healthcare is in the midst of a data revolution. Years of retaining vast data sets of comprehensive patient records is coming to fruition with AI, our doorway to superior healthcare capabilities.
 McKinnely, S. M. (2020) International evaluation of an AI system for breast cancer screening. Nature, 577; 89-94
 Boland, G. W. L. (2009) The radiologists conundrum: benefit and costs of increasing CT capacity and utilization. European Radiology, 18; 9-11.
 Mcdonald, R. J. (2015) Thee effects of changes in utilization and technological advancements of cross-sectional imaging on radiologist workload. Academic Radiology, 22; 1191-1198
By Jai Bains, Oncology Market Forecast Analyst.