Current status and future direction of cancer research using artificial intelligence for clinical application

1 INTRODUCTION

The expectations for AI are rising due to rapid progress in machine learning brought about by the advent of deep learning, the development of high-performance GPUs and other computer analysis infrastructure, and the ability to utilize big data through the development and expansion of public databases.^1-3 Currently, we are in what is being referred to as the third AI boom,⁴ defined by the broad social implementation of AI technology. AI technology is widely used in today's society, from the security field (e.g., facial recognition at airports) to automatic translation, automatic driving, and general consumer electronics.^5-7 The medical field is no exception, with more than 800 AI-based medical devices approved by the US FDA.⁸ In Japan, a number of AI-equipped medical devices, including those developed by our group, have obtained regulatory approval and are already in clinical use. Since transformer was developed in 2017,⁹ generative AI research, as represented by ChatGPT, has undergone rapid development, with ever-increasing expectations.^{10, 11} Some experts have proposed that we are already in the fourth AI boom.¹² Following the LLM, LMM research has gained momentum, with Med-PaLM as an example from the medical field. Generative AI-driven medical research has been continuously published.^{13, 14}

In cancer research, AI is actively explored for medical image analysis, such as an endoscope, radiation, and pathology image-based diagnostics, with advances already applied clinically.^15-21 In addition to medical image analysis, AI is also being used in various cancer research applications, including omics analysis, medical information analysis, electronic medical records using NLP, and the analysis of pathology and radiology reports.^22-27 Further, in 2003, the whole genome analysis project, which had been analyzed on the basis of an international consortium, was completed, and the world entered what is now known as the postgenomic era. The applications of genomic information in medicine have expanded, and the concept of personalized medicine emerged.²⁸ In particular, the announcement of the Precision Medicine Initiative by US President Barack Obama in January 2015 had a major impact on global healthcare policy.²⁹ This is especially true in the field of cancer research, where the concept of precision oncology is actively pursued. In this context, AI is increasingly used to optimize cancer treatment selection based on genomic and medical information.^{30, 31}

While the expectations for AI in cancer research become ever greater, challenges remain. Examples include overfitting, wherein high accuracy is observed in training data but declines in test data; the black box problem, where AI analysis is so complex that humans (medical professionals) cannot understand it; and the domain shift problem, which arises due to differences in medical equipment and protocols between institutions. In order to advance AI-based cancer research and achieve optimal results, these issues should be considered.

In this review, the history of medical AI is first described, followed by an introduction to the current status of AI applications in cancer research, with a discussion on issues and future directions in AI-based cancer research.

2 BRIEF HISTORY OF MEDICAL AI

The concept of AI itself can be traced back to an anecdote in ancient mythology in which a divine master craftsman bestowed intelligence and consciousness onto an artifact.³² The coining of the term “artificial intelligence” by John McCarthy and colleagues at the Dartmouth Conference of 1956 is considered the birth of modern AI.³³

A key event in the foundation of medical AI is considered to be the initiation of research into CAD. These early CAD systems used flowcharts, statistical pattern matching, probability theory, and knowledge bases to drive the decision-making process.³⁴ In the 1970s, expert systems, an AI technology, attracted significant attention worldwide, and the MYCIN expert system, a very early CAD system and recognized as the world's first successful medical AI system, was developed (Figure 1).^35-37 MYCIN, a backward-linked expert system, identified bacteria that cause serious infections such as bacteremia and meningitis, recommending antibiotics with dosage adjusted to the patient's weight. It uses a simple inference engine and works with a knowledge base consisting of about 500 rules. Physicians were asked a series of simple yes/no or text questions and provided a list of causative organisms ranked from high to low based on the probability of each diagnosis, their level of confidence in the probability of each diagnosis, the reason behind each diagnosis, and the recommended course of drug treatment. A study conducted at Stanford University School of Medicine showed a 65% correct response rate for bacterial infection diagnosis by MYCIN, which was not comparable with that of specialist physicians, but was higher than that (42.5%–62.5%; average 55%) of five non-specialist Faculty member clinicians.³⁷ However, MYCIN was never widely used in clinical practice, as there were concerns about liability issues upon misdiagnosis, and professionals were reluctant to accept the results produced via CAD. Furthermore, expert systems were often unable to formulate expert knowledge, especially with exceptions arising among the rules.³⁸ The shortcomings of expert systems were subject to discussion, which brought about a downturn in AI research (the second AI winter period). Although MYCIN has never been used in clinical practice, it is known worldwide as a pioneering medical AI system. Despite the fact that it showed higher accuracy than general physicians, although it did not reach the level of specialist physicians, the results of MYCIN have influenced the development of diagnostic and decision support tools using current AI technology, including those used in the field of oncology.

Since then, CAD has remained actively studied, with the first commercial CAD system for mammography, the ImageChecker M1000 system, approved by the US FDA in 1998.³⁹ Subsequently, CAD for detecting lung cancer (nodule shadows) in simple chest radiographs and chest CT images and CAD for detecting polyps in colon CT examinations were successively approved by the US FDA.⁴⁰ Insurance coverage for the use of mammography CAD became available in the United States in April 2001 (Medicare coverage), a major factor in spurring the spread of CAD.⁴¹

In 2006, auto-encoder, a deep learning technology, was developed. This advent of deep learning was the main catalyst for the emergence of AI technology, which ushered in what is generally referred to as the third AI boom. In 2015, a group at Microsoft demonstrated recognition accuracy that exceeded the average human error rate in ILSVRC by using deep learning techniques.⁴² In 2017, AlphaGo, an AI-based Go program, defeated the world's top Go player.⁴³ Thus, reports of AI surpassing human capabilities one after another invigorated AI research and development in various fields in society, including the medical field, where it has been centered on medical image analysis. Importantly, results are not only at the basic research level, but AI-powered software as a medical device (AI-SaMD) is also being applied clinically after receiving medical device approval. In response to this situation, an independent organization has published AI-based software (CE certified in the EU) that can be used in clinical radiology practice, with the aim of increasing the transparency of AI in radiology (Health AI Register: https://radiology.healthairegister.com).⁴⁴ The publicly available information is summarized in Figure 2 based on the type of medical imaging exam and the body part. Most approvals for medical imaging exams are for X-ray, CT, and MRI image analysis, and there are many approvals for AI-SaMD that target the neuro region and chest.

3 AI TECHNOLOGIES USED IN CANCER RESEARCH

A major factor that has led to a significant focus on AI in recent years has been the emergence of deep learning, and since deep learning is also a machine learning method, it can be concluded that machine learning is used as the underlying technology for current AI. In Figure 3, we present the properties of machine learning and examples of its use in cancer research.

Machine learning can be divided into four types: supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. Supervised learning corresponds to regression/classification tasks, unsupervised learning to clustering and dimensionality compression tasks, semi-supervised learning to tasks with generative or discriminative models, and reinforcement learning to optimization tasks. It should be emphasized that current AI is machine learning based and thus limited to the above tasks. There is a tendency to place excessive expectations on AI due to mass media and other factors; however, there are limits to the tasks that AI can actually handle, and it is thus necessary to understand its characteristics.

The main methods used and examples in cancer research are shown in Figure 3. Note that these methods are not limited to a single task (e.g., SVM is used for regression tasks in addition to classification tasks). Extensive work has been done in cancer research, particularly in supervised learning classification tasks, yielding results that have been applied clinically. This is attributed to the proficiency of deep learning at image analysis, with many researchers working on classification (benign or malignant, etc.) tasks, and also to supervised learning based on annotation by medical specialists being important to apply for the approval of AI-equipped medical devices.^45-49

More recently, generative models have received considerable attention in cancer research. For example, Das et al. proposed the NAS-SGAN model for automated breast cancer classification using quantitative analysis of histopathological images to improve classification model performance with limited annotation data.⁵⁰ The model is a semi-supervised learning framework combined with deep neural network-based generative adversarial learning. Even a limited amount of labeled samples could better discriminate between tumor grade, thereby improving system robustness and accuracy. Pang et al. developed a semi-supervised GAN-based radiomics model for augmenting breast ultrasound images.⁵¹ The model produced high-quality breast ultrasound images, validated by two experienced radiologists, with semi-supervised learning improving the quality of generated synthetic data compared with the baseline method. More accurate breast mass classification was thus achieved compared with other state-of-the-art methods.

4 CURRENT STATUS OF AI-BASED CANCER RESEARCH

In this section, we introduce the current state of cancer research using AI, focusing mainly on the condition in Japan. Table 1 shows a list of AI SaMDs related to oncology that have been approved for use in Japan; they are all products related to medical image analysis. The classification of classes here is based on the international system that has been adopted in Pharmaceutical Affairs Law in Japan, with Class II defined as “those that are considered to have a relatively low risk to the human body even if a problem occurs” and Class III defined as “those that are considered to have a relatively high risk to the human body if a problem occurs.” One of the characteristics of AI SaMDs approved in Japan is that many of them are AI systems that support endoscopic diagnosis. Of the 18 types of AI-SaMD approved for tumor-related applications, nine (50%) are for the purpose of supporting endoscopic diagnosis. At the research level, in addition to image analysis, there is also active research on omics analysis, NLP, and analysis using LLMs. Here, we focus on three areas: 1. medical image analysis, 2. omics analysis, and 3. NLP and LLMs.

4.3 Cancer research using NLP and LLMs

NLP, one elemental technology of AI, has been reported to extract important clinical variables from free text in electronic medical records to identify cancer cases, quantify staging and treatment results, and automatically extract information on stage, histological type, tumor grade, treatment method, etc. with high accuracy from clinical notes, pathology reports, and surgical reports of patients with lung cancer.^{80, 81} Nara Institute of Science and Technology and Osaka International Cancer Center are conducting automatic analysis using NLP technology on spoken language to evaluate the effects of cancer treatment on the cognitive abilities of cancer patients.⁸² This study aims to assess cognitive function in cancer patients in a less burdensome manner.

An important point in the NLP field is the rapid development of LLMs with the advent of transformers, and their use in various aspects of society, as exemplified by ChatGPT, has had a significant impact. The introduction of LLMs into the medical field is also underway, with GPT-4 reportedly achieving passing scores on Steps 1–3 of the USMLE by more than 20 points, making it the first computer-based system to qualify for standardized US physician face-private examination.⁸³ Conversely, despite its potential to provide medical information, GPT-4's training data are not based on expert medical knowledge. Recent studies have reported that ChatGPT may obtain fundamentally flawed or inadequate information in the field of urologic oncology due to the extrapolation of data from relevant literature and abstracts without logic or accuracy.⁸⁴ Therefore, current LLM-based chatbots may not be reliable enough for direct use. To address these challenges, it has been reported that the integration of the EAU oncology guidelines into ChatGPT-4 has the potential to address questions that were previously difficult to answer accurately.⁸⁴ Fine tuning of LLMs is also underway to adapt these to specific medical reasoning tasks. Therefore, we would like to introduce the results of a study presented by a group from the Mayo Clinic, in which LLMs were retrained to be adapted to specific diseases, with favorable results. Zhu et al. at the Mayo Clinic developed an LLM retrained to meet the needs of patients with HN cancer (HN) undergoing radiation therapy, with an emphasis on symptom management and cybership care (Figure 4).⁸³ Initially, they built a comprehensive external database to “educate” ChatGPT-4, integrating expert consensus guidelines on supportive care for HN patients and correspondence on 90 HN patients from physicians and nurses in the Mayo Clinic electronic health record. Performance was assessed using 20 patient post-treatment inquiries, and evaluated by three radiation oncologists (RadOncs). The custom-trained model was highly accurate in assisting HN patients with evidence-based information and guidance on symptom management and survivorship care. In contrast, the system introduced here uses a database from the United States, but when we consider factors such as differences in disease and treatment effects by race, medical systems, and language, the ideal method to build a more accurate system may be to build a similar system for each region and country (in the case of Japan, a Japanese system based on a Japanese database). In order to develop cancer research using NLP and LLMs, it is particularly important to construct high-quality large-scale medical databases. As medical information is sensitive personal information, care must be taken when handling it; however, by building a high-quality large-scale medical database, this field should develop constructively.

5 CHALLENGES IN CONDUCTING AI-BASED CANCER RESEARCH

Here we presented the current status of AI-based cancer research for clinical application. While the main focus was on successful AI-based cancer research, there are recurring challenges specific to medical AI.^{15, 77} The first is “overfitting,” a condition in which the training error is small but the generalization error (the error when judging unknown data) is not. In the medical field, where the number of supervised data is limited, carefully judging the generalization performance of the constructed training unit is essential. The second is the “black box problem,” which refers to a situation in which the analysis process by AI is so complex that humans cannot understand it. This highlights the need to develop XAI/interpretable AI. The third issue is the “domain shift problem,” in which characteristics such as differences in the manufacturer, model number, and year of the medical equipment used, as well as differences in protocols at different facilities, can be stronger than the characteristics of the pathology, significantly affecting the robustness of the AI built as a result. The fourth issue is the “hallucination problem,” in which AI generates incorrect information, and should be noted in studies that utilize LLM.⁸⁵ Although we omit detailed descriptions in this review for reasons of space limitation, various countermeasures are being taken by us and others against the “overfitting,” “black box,” “domain shift,” and “hallucination” problems.^{15, 71, 85, 86} In addition to technical issues, there are also related to legal systems and guidelines. Medical AI research and development involves the large-scale use of medical data, which is often sensitive personal information, resulting in many sensitive points in terms of privacy protection. In particular, as there are more opportunities to conduct cloud-based research when utilizing LLM, the utmost care must be taken to prevent the leakage of personal information. Furthermore, all medical devices using AI in clinical practice so far are “locked” AI medical devices with fixed algorithms that have been approved. To take full advantage of AI functions, “adaptive” AI medical devices that can learn and change through use must also be approved. The revised Pharmaceutical Affairs Law, which went into effect in Japan on September 1, 2020, introduced IDATEN (Improvement Design within Approval for Timely Evaluation and Notice) as an approval system for medical devices that appropriately considers the characteristics of medical devices that are continuously improved and enhanced as well as technological innovations such as AI.⁸⁷ It is thus expected that “adaptive” AI medical devices will be approved in the near future.

6 CONCLUSION

Judging from its technological potential, there is no doubt that AI technology will be used in various aspects of cancer research in the future. Conversely, as mentioned above, there are multiple issues that need to be addressed. Thus, we believe that AI-based cancer research and its clinical application should not be rushed. Rather, considerable caution should be practiced. We would like to emphasize the need to rigorously verify the clinical utility of AI products by conducting appropriate clinical performance evaluation studies. Priority should be placed not on the interests of individual researchers, but on the benefits to patients and the public, and this cannot be pursued through short-sighted strategies. Thus, we believe that a common effort should be mobilized to advance the utilization of AI in cancer research, driven not only by experts in medicine and informatics, but also by experts in law, bioethics, sociology, and patient associations, aiming to benefit society through repeated discussion.

AUTHOR CONTRIBUTIONS

Ryuji Hamamoto: Conceptualization; funding acquisition; investigation; project administration; visualization; writing – original draft; writing – review and editing. Masaaki Komatsu: Conceptualization; investigation; writing – review and editing. Masayoshi Yamada: Conceptualization; investigation; writing – review and editing. Kazuma Kobayashi: Conceptualization; investigation; writing – review and editing. Masamichi Takahashi: Conceptualization; investigation; writing – review and editing. Mototaka Miyake: Conceptualization; investigation; writing – review and editing. Shunichi Jinnai: Conceptualization; investigation; writing – review and editing. Takafumi Koyama: Conceptualization; investigation; writing – review and editing. Nobuji Kouno: Conceptualization; investigation; writing – review and editing. Hidenori Machino: Conceptualization; investigation; writing – review and editing. Satoshi Takahashi: Conceptualization; investigation; writing – review and editing. Ken Asada: Conceptualization; investigation; writing – review and editing. Naonori Ueda: Supervision; writing – review and editing. Syuzo Kaneko: Conceptualization; investigation; writing – review and editing.

FUNDING INFORMATION

This work was supported by the Cabinet Office BRIDGE (programs for bridging the gap between R and D and the ideal society (Society 5.0) and generating economic and social value) and the RIKEN Center for the Advanced Intelligence Project.

CONFLICT OF INTEREST STATEMENT

Dr. Ryuji Hamamoto is an editorial board member of Cancer Science. R.H. and M.K. have received research grants from Fujitsu Limited, and R.H. and K.K. have received research grants from the Fujifilm Corporation. The other authors declare no conflicts of interest.

ETHICS STATEMENT

Approval of the research protocol by an Institutional Reviewer Board: N/A.

Informed Consent: N/A.

Registry and the Registration No. of the study/trial: N/A.

Animal Studies: N/A.