Artificial Intelligence in Oncology: Current Applications and What the Future Holds

Artificial intelligence (AI) has been used in oncology for many years. Recent studies, however, show that AI may offer numerous additional applications in cancer diagnosis and treatment. Over the past few years, AI applications in oncology have moved from a promising future into reality, and ongoing research continues to reveal many further potential uses in cancer care.

The AI applications currently used in oncology practice mainly assist with cancer diagnosis . Machine learning and deep learning models are used to improve the accuracy and efficiency of the cancer diagnostic process.

The scope of the concepts "artificial intelligence," "machine learning" and "deep learning" is illustrated below.

Artificial Intelligence Today

The US Food and Drug Administration (FDA) has approved numerous AI technologies for use in oncology, with radiology accounting for some of the most notable applications.

According to a 2021 analysis, AI-associated devices approved for oncology use are most commonly used in

• radiology (54.9%) and
• pathology (19.7%), and
• breast cancer is their most frequent application (31%).

Breast radiologists increasingly use AI to

• determine breast density,
• assess mammogram quality,
• classify mammograms as indicating low, intermediate or high breast cancer risk,
• identify breast cancers
• and also  identify the risk of atherosclerotic disease (cardiovascular blockages) in patients.

AI can also provide additional information about whether a finding on breast ultrasound is cancerous.

AI has also proved effective in colorectal cancer screening and diagnosis. In 2021, the FDA authorized marketing of GI Genius, the first device to use AI to detect lesions during colonoscopy.

Some pathology laboratories also use AI to read digital pathology slides and clarify cancer diagnoses, following years of research in this field.

The applications discussed so far have all involved image recognition . However, AI is also beginning to be used in the development of new cancer drugs .

Some pharmaceutical companies use AI in oncology to discover new, safe molecular targets , while others sometimes combine AI with molecular modeling to design entirely new drug candidate molecules . The future of AI in drug discovery is very promising.

AI is also used to improve the efficiency of radiotherapy planning and define tumor and organ contours, thereby speeding up treatment and improving the effectiveness and safety of radiation delivery.

In supportive or palliative care, artificial intelligence (AI) is used to assess vital signs self-reported by remotely monitored patients . Natural language processing (NLP) and machine learning (ML) models can identify adverse health trends at an early stage and issue the necessary alerts promptly. These models can automatically detect critical changes in health status and provide information for rapid intervention, potentially improving the quality of patient care.

Artificial Intelligence in the Future

Many ongoing studies are revealing the potential for additional AI applications in oncology.

Breast imaging is considered a fertile field for AI research. Future applications, used independently of or together with family history, may become important tools for determining breast cancer risk over both the short and long term.

Researchers are examining whether AI can use abdominal imaging and longitudinal electronic health records to identify people at high risk of pancreatic cancer . By enabling earlier detection of a disease associated with high mortality, this could make it possible to establish screening programs for high-risk individuals.

AI may eventually be able to assess the molecular characteristics of cancers and their biomarkers , potentially eliminating the need for invasive biopsies and the wait for results in some situations.

This could be achieved through a noninvasive approach using a blood sample or MRI (magnetic resonance imaging) radiomics , known as a "virtual biopsy". For example, a deep learning model could analyze a brain MRI to determine whether a brain tumor has a specific genetic mutation, such as IDH1 or BRAF.

A New Era in Artificial Intelligence with ChatGPT

As a major new step, large language models such as ChatGPT are emerging as an exciting development with potential applications in oncology.

Studies indicate that current ChatGPT models can answer medical questions with good accuracy . However, these findings need more systematic validation using high-quality, real-world datasets.

Hospitals and medical centers may soon be able to train high-quality large language models (LLMs) using information such as laboratory results, scans and previous medical histories from petabytes of medical oncology data . Access to such data could significantly advance AI applications in oncology. It could enable the development of more effective and personalized AI models for cancer diagnosis, treatment and research, thereby improving patient care and treatment outcomes.

Examples of ChatGPT in Oncology

In a study published earlier this year, Haver and colleagues found that ChatGPT answered breast cancer screening and prevention questions, as assessed by breast radiologists, with 88% accuracy .

Two studies published in August 2023 showed that chatbots can answer questions about various cancers with high accuracy, while also revealing continuing limitations of the technology.

In one of these studies, Pan and colleagues evaluated chatbot responses to the most frequently searched online questions about skin, colorectal, prostate, lung and breast cancers. The chatbots generally provided high-quality information, but it was not always actionable and was written at a university reading level.

In the other study, Chen and colleagues found that a chatbot's responses about cancer treatments were not always consistent with recommendations in US National Comprehensive Cancer Network (US-NCCN) guidelines. The chatbot could provide at least one treatment recommendation for 98% of questions about breast, prostate and lung cancer, and all these responses included at least one US-NCCN-consistent recommendation. However, 34.3% of responses included at least one inconsistent recommendation, while 12.5% were "hallucinated," meaning they were not part of any recommended treatment.

Another application of AI is to predict survival outcomes in people with cancer using natural language processing models. Nunez and colleagues documented that these models could predict survival outcomes using only data from a patient's initial oncology consultation.

Optimizing AI for Clinical Use

Experts note that several limitations and issues must be addressed to optimize the use of artificial intelligence (AI) in clinical oncology.

From an ethical perspective, AI models and AI-enabled medical devices should be developed using diverse datasets and should accurately represent the patient groups they will serve in clinical practice. Otherwise, these models and devices may amplify bias and healthcare disparities.

During research and development (R&D), one barrier to using large, diverse datasets is medical centers' reluctance to share data.

We need to normalize data sharing and find ways to encourage it, ideally with patients' support. Alternatively, we must develop new training algorithms that train AI models locally, combine models across different centers and then validate the accuracy of the combined models.

Despite all these advances, many current models still require evaluation in clinical practice.

An AI model sitting on a shelf is merely a model. In real-world oncology, it should be used in clinical care, provide value to physicians and improve outcomes for patients and health systems. The usability, acceptance and safety of these models must be established, and ongoing evaluation is needed to understand their impact.

The adoption of AI models in clinical settings is currently limited by implementation challenges , including

• difficulties integrating AI models into electronic health record software,
• a lack of intuitive user interfaces in many AI models, and
• the high costs of training some of the latest models.