Beyond the Triad: Integrating 'Knowledge' into Medical AI—Emerging Challenges and New Paradigms

Although it is a well-worn topic, artificial intelligence (AI) ultimately cannot bypass the three core elements: algorithms, computing power, and data. These elements are closely interlinked, holding the key to the lifeblood of AI technology. As AI-powered products in healthcare settings become increasingly diverse, their advancement undoubtedly relies on the rapid development of these three pillars. To further explore the technological advancements and future industrial trends of medical AI, top experts in the field offered new insights into this triad at the recently held Global Conference on Artificial Intelligence and Robotics (CCF-GAIR 2020). The panel included Zhou Zhihua, Chair of the Department of Computer Science at Nanjing University; Gong Qiyong, Vice President of West China Hospital; Zheng Yefeng, Director of Tencent’s Tianyan Laboratory; Xie Guotong, Chief Medical Scientist of Ping An Group; and Shen Dinggang, Co-CEO of United Imaging Intelligence.

Gong Qiyong, Vice President of West China Hospital and a radiologist, believes that AI intervention is indispensable for addressing human neurological disorders.

In Gong Qiyong’s research on brain MRI scans of patients with schizophrenia, he discovered that disease progression drives structural changes in the brain, causing multiple brain regions to either expand or shrink. If such changes can be quantified through imaging, clinicians may be able to perform detailed classification of neuropsychiatric disorders, including schizophrenia, depression, and obsessive-compulsive disorder.

At the summit, he cited a previous research finding: “A physician once observed that when one of her patients experienced obsessive-compulsive symptoms, electrical activity in a specific brain region became abnormally heightened. She therefore attempted to stimulate that area with electrodes, resulting in a significant improvement in the patient’s condition.”

Unfortunately, these results did not demonstrate definitive efficacy in subsequent large-scale experiments. In this regard, Gong Qiyong stated, “Given the highly complex structure of the human brain, it is extremely challenging to accurately localize cerebral lesions using two-dimensional imaging. We cannot determine whether physicians precisely stimulated the marked regions during procedures. Therefore, we need AI technology to rapidly and accurately model the brain to assist in localization.”

“If we can accurately obtain patients’ brain information and lesion status through AI, subsequent therapeutic research may also benefit from this.”

Dean Gong Qiyong’s research appears to herald that AI can indeed facilitate the advancement of neuroradiology and promote the diagnosis and treatment of neurological disorders. In practice, however, the paucity of data undoubtedly constitutes a major obstacle impeding research progress.

Generally, due to high costs, patients with suspected psychiatric disorders do not undergo MRI scans directly during examination. Instead, physicians can rapidly complete an initial diagnosis based on the patient’s clinical presentation and carefully designed rating scales, thereby proceeding directly to standard treatment protocols. This practice makes it difficult to collect brain MRI images from psychiatric patients.

Dean Gong Qiyong also proposed multi-center collaborative clinical studies to address data issues. By pooling meticulously annotated brain imaging data from multiple hospitals, the sample size can indeed be expanded; however, it remains challenging to achieve the large-scale datasets desired for AI training.

At this juncture, several novel approaches have emerged.

Few-shot learning is a new direction driving the development of deep learning and one of the key research focuses at Tencent’s Tianyan Laboratory. Zheng Yefeng, Director of Tencent Tianyan Laboratory, introduced two few-shot learning algorithms at the conference, aimed at addressing the issue of insufficient medical imaging data in the future.

The first approach is transfer learning, which involves transferring a model trained on one task (the source domain) to another task (the target domain). In this way, we can leverage the abundant data from the source domain to assist learning with limited samples in the target domain, specifically manifested in the “pre-training + fine-tuning” paradigm.

Zheng Yefeng illustrated the principle of transfer learning with an engaging example: “Suppose you want to undertake a computer vision project for ‘tiger recognition,’ but you lack frequent access to observe tigers. However, if you could keep an orange tabby cat, you could capture certain features from its daily behaviors—such as eating, playing, and resting—through close, day-to-day interaction, and then apply these insights to tiger recognition.”

The second approach is unsupervised domain adaptation learning, designed to address the issue of domain shift. In this area, Tencent’s Tianyan Laboratory has proposed a novel unsupervised domain adaptation (UDA) method to mitigate performance degradation caused by domain shift. This method requires only source domain data and annotations, along with a subset of images from the target domain, achieving adaptation between the two domains without the need for new annotations.

However, can merely adjusting algorithms solve the problems in medical artificial intelligence? In fact, there is still room for optimizing the algorithms themselves.

Zhou Zhihua, Chair of the Department of Computer Science at Nanjing University, cited GPT-3, currently the largest artificial intelligence model, as an example to discuss the “aesthetics of brute force” in contemporary machine learning. This model was trained on 45 TB of data, comprises 175 billion parameters, and has a total size of 700 GB. Based on this model, significant progress has been made in addressing many challenging natural language processing tasks.

From this perspective, the expansion of models and data can indeed effectively improve model quality; however, medical AI imaging also faces data-related challenges.

Assuming physicians are able to annotate a sufficient volume of medical images, who will bear the exorbitant computational costs? A single training run for GPT-3 consumes computing power valued at approximately $13 million. Consequently, even if bugs exist within the model, no party is willing to modify them and incur the substantial expense of re-running the training process.

Regarding this issue, Zhou Zhihua and Zheng Yefeng proposed different approaches.

In his speech, Zhou Zhihua stated, “Such models can solve many problems and deliver significant performance improvements. However, the prohibitively high costs also pose new challenges for those of us engaged in artificial intelligence research. This warrants particular reflection by academics in the field: costly investments must yield commensurate returns. In industry, enhanced model performance translates into economic benefits; sometimes, even a one- or two-percentage-point improvement can generate returns sufficient to offset the investment. But if academia incurs such substantial costs, how can it ensure adequate returns?”

“Leave the pursuit of ‘performance’ to industry; academia need not focus excessively on ‘performance,’ as marginal gains in model performance hold little significance for the academic community. Merely publishing a few papers does not justify such substantial investment. So, what should academia do? Return to first principles and fulfill its core mission: blazing trails and envisioning the future.”

In contrast, Zheng Yefeng told VCBeat in an interview, “A 1% efficiency gain holds little significance for enterprises, whereas academia should strive for excellence and pursue more perfect outcomes.”

Which Path Is Better? It is too early to draw a definitive conclusion, but judging from the development of AI, collaboration between industry and academia may be the most promising direction at present.

Nearly 70 years have passed since the birth of artificial intelligence, yet researchers have not broken free from the influence of the three core elements. However, as the technology gradually matures, Zhou Zhihua argues that “knowledge” must be introduced as an essential element in the pursuit of next-generation artificial intelligence.

Integrating “machine learning” with “logical reasoning” is the “holy grail” problem in the field of artificial intelligence. Previous efforts have either emphasized “reasoning” or “learning,” but favoring either side fails to fully harness the power of AI. Zhi-Hua Zhou has proposed “abductive learning,” aiming to enable machine learning and logical reasoning to function more balancedly and effectively within a unified framework.

“From the current perspective, there are two main approaches to combining ‘machine learning’ with ‘logical reasoning.’ In the first approach, machine learning is introduced into logical reasoning, but the core problem-solving process still relies on reasoning; thus, we describe it as ‘reasoning-heavy and learning-light.’ The second approach is essentially the reverse: logical reasoning techniques are integrated into machine learning, but the primary problems are ultimately solved by machine learning, so we term it ‘learning-heavy and reasoning-light.’ In both cases, the imbalance implies that the potential of one side’s technology is not fully leveraged.” In contrast, “abductive learning” represents a true fusion of the two.

Zheng Yefeng also expressed strong agreement with Zhou Zhihua’s views, but he highlighted the implementation challenges: “In the early stages of artificial intelligence, we attempted to integrate prior knowledge into AI algorithmic models. However, different types of knowledge require distinct integration methods, and there is no universal approach. Furthermore, our models at that time were less effective than black-box models supported by large volumes of data, which was somewhat discouraging. Nevertheless, times have changed, and the incorporation of knowledge may now fundamentally transform current artificial intelligence.”

Whether it be algorithms, computing power, data, or potentially integrated knowledge, everything will serve physicians and specific clinical scenarios; only real-world scenarios are the true litmus test for AI.

In this regard, Xie Guotong, Chief Medical Scientist at Ping An Group, believes that the development of AI must leverage the power of its ecosystem, and the payment for medical technology is a three-way contest among enterprises, health insurance providers, and patients.

He stated in the interview, “The National Health Commission places great trust in us because we have an insurance ecosystem. Our incentive is to keep patients healthy and long-lived to control insurance risks. Therefore, our interests are aligned with those of the National Health Commission.”

He also stated, “In fact, medical imaging is just a small department within the healthcare system. The application of AI in this field is highly effective and represents an inevitable trend. It is normal for some peers to fall along the way. Even mistaken experiences are valuable; all such accumulation will help us go further.”