Is Deep Learning in Healthcare Reliable? An Analysis of Its Viability and Recent Advances

Many AI researchers believe that the next major application for deep learning lies in healthcare. Hold on a moment. While its precise judgments may indeed align with the medical field’s pursuit of accuracy, to what extent has the healthcare sector actually embraced this technology?

In many technology-driven industries, continuous technological breakthroughs are essential for survival. However, other sectors, constrained by regulations and other factors, tend to adhere to traditional practices. The healthcare industry falls into the latter category. Despite the significant progress brought about by the implementation of precision medicine initiatives and the contemporary wave of mobile health, healthcare, as a vital component of the national economy, remains more conservative under various constraints, exhibiting lower acceptance of potentially advanced tools.

Deep learning is no longer a novel topic, and its application scenarios continue to expand widely. Traditional deep learning techniques are likely to evolve into more mature variants—convolutional and recurrent neural networks being their initial forms—providing specialized functionalities for various medical fields and unlocking greater potential. Of course, we will soon face a series of challenges as well.

Over the past year, deep learning has shifted its focus from commercial domains such as video, medical imaging, and speech recognition analysis to scientific research. In healthcare, deep learning will become more deeply embedded inMedical Imaging, Sensor-Based Data Analysis, Translational Bioinformatics, Public Health Policy Developmentand other aspects.

A team of professional deep learning researchers conducted a specialized analysis of the applications of deep learning in healthcare informatics in recent years. They pointed out that advancements in computing power, rapid data storage, and parallel computing have laid the foundation for the rapid development of deep learning, while its predictive capabilities and automatic recognition functions have made it widely popular in disease diagnosis. Additionally, they found that the adoption of deep learning technology in the healthcare industry has not only increased significantly in frequency but also undergone certain changes in variety. As the most predominant type, the usage share of Recurrent Neural Networks (RNNs) decreased from nearly 90% in 2010 to 40% in 2015, whereas the usage frequency of other improved variants has been steadily rising.

In the field of medical informatics, the ability of deep learning to automatically discover new features is highly valuable. For example, without human intervention, deep learning algorithms can identify numerousContrast features of extremely high complexity, difficult to describe exhaustively in words. These subtle features may be indicative of fibroids or polyps. Researchers also pointed out that in the field of public health, deep learning can identify macro-level patterns from complex regional and demographic data.

“Most of the features tracked by deep learning models are difficult to interpret,” the researchers pointed out. “Therefore, users typically treat deep learning as a ‘black box’ approach, deeming it neither possible nor necessary to explain how it arrives at correct judgments.” However, the problem with this black box nature of deep learning is that neural networks can be “deceived” during the training process.

So-called “adversarial attacks” involve introducing minor perturbations to input data, which can easily mislead deep learning systems. For instance, adding subtle noise to a medical image may cause the sample to be incorrectly excluded from its classification; conversely, meaningless synthetic samples are sometimes erroneously assigned to a specific category. This represents a genuine limitation of deep learning technologies. Of course, this issue is not unique to deep learning; any machine learning approach may encounter similar problems.

Training any type of neural network to recognize simple features requires large amounts of data. Moreover, this technologyNot applicable to all cases, especially rare diseases. For deep learning, overfitting remains a challenge, making it difficult for neural networks to achieve generalization.

Although deep learning holds great promise in healthcare, it faces skepticism: What proportion of medical applications can this technology truly occupy? Is it really necessary to introduce deep learning in certain fields? While it is touted as a replacement for physicians’ manual judgments with higher accuracy, can it genuinely ensure improved efficiency in practice? We still lack answers to these questions. However, one thing is certain: many large companies are increasingly leveraging deep learning to explore the healthcare landscape, and healthcare startups mastering this technology are emerging in rapid succession.

1. IBM: Using Deep Learning to Identify Mitosis in Cancerous Cells

When diagnosing cancerous cells, biopsy is typically used to analyze patient tissue samples. During sample analysis, typical tissue samples are stained with reagent solutions; the intensity of the reagent color and its distribution within the cellular tissue indicate the type of disease and its severity.

However, these tissues are sometimes exceedingly minute, requiring medical experts to employ alternatives to naked-eye examination to detect key features indicative of tumor cell clearance or malignant transformation, thereby facilitating subsequent clinical decision-making. Notably, at the 2016 MICCAI International Conference’s “Tumor Proliferation Assessment Challenge,” researchers from IBM Laboratory achieved commendable results by leveraging artificial intelligence to characterize tissue samples.

2. Google DeepMind: Deep Learning for Medical Records, Eye Diseases, and Cancer Treatment

In February last year, DeepMind established the DeepMind Health division and acquired Hark, a company specializing in healthcare management applications, leveraging its deep learning expertise to address the shortcomings of traditional paper-based medical records.

Last July, in collaboration with Moorfields Eye Hospital, we developed a deep learning system to identify visual disorders, aiming to detect early signs of eye diseases such as age-related macular degeneration and diabetic retinopathy, thereby enabling early prevention of visual impairments.

Last August, DeepMind also used deep learning algorithms to design radiotherapy regimens for patients with head and neck cancer, shortening treatment time and reducing radiation-induced damage.

3. NVIDIA: Cancer Distributed Learning Environment Initiative

NVIDIA, a computer graphics chip manufacturer determined to delve deeply into deep learning, announced last year that it would collaborate with the U.S. National Cancer Institute and the U.S. Department of Energy to develop an artificial intelligence computing framework to assist in cancer research. The framework is named the “Cancer Distributed Learning Environment” (CANDLE).

There are hundreds of types of cancer, each with thousands of potential causes, making the selection of appropriate therapies a formidable challenge. The CANDLE initiative leverages deep learning algorithms to identify patterns and trends within vast amounts of healthcare data, assisting researchers in predicting how specific tumors will respond to particular drugs and elucidating the factors driving cancer cell proliferation.