AI as an Aid, Not a Replacement: Enhancing Medical Practice without Supplanting Doctors

In this blog, I emphasize the crucial point that AI is not a replacement for doctors. Rather, its role is to unburden doctors from the tedious tasks of keyboard input and time-consuming procedures, thereby enhancing their capabilities and allowing them to focus on what truly matters: patient care.

The foremost strength of deep learning resides in its remarkable computational capacity to swiftly integrate and analyze data derived from the diagnostic decisions of numerous doctors. Essentially, deep learning models serve as rapid calculators, excelling in processing and synthesizing vast amounts of information. However, they encounter a significant limitation: unlike humans, these models struggle to adapt to new and evolving situations with the same intuition and flexibility. While recent advancements in generative models have showcased their ability to produce novel outputs, these innovations often amount to recombination of existing data. Deep learning inherently lacks the human-like ability to genuinely innovate or understand the complexities of new scenarios.  On the flip side, AI excels in quickly analyzing and integrating the value of existing knowledge by performing calculations on a scale far beyond human capacity, leading to the generation of superior recommendations that surpass human expectations.

Understanding the potential and limitations of Deep Learning (DL) is crucial for directing the course of future research. The fundamental objective of DL is to learn a mapping function $f: \mathbf{x} \mapsto \mathbf{y}$, where $\mathbf{x}$ represents the input (for example, a medical image) and $\mathbf{y}$ denotes the desired output. This mapping function $f$ is represented by a neural network, which is trained using a paired dataset $\{(\mathbf{x}^{(1)}, \mathbf{y}^{(1)}), \cdots, (\mathbf{x}^{(N)}, \mathbf{y}^{(N)})\}$ (for supervised learning). The learning process of $f$ involves the use of backpropagation to minimize a loss function, thereby adjusting the neural network to approximate the actual output $\mathbf{y}^{(k)}$ as closely as possible to the predicted output $f(\mathbf{x}^{(k)})$ for each training instance $k = 1, \cdots, N$.  The success of the training process heavily relies on the quality of the training data, which in turn depends on the expertise of the medical professionals who annotate and curate this data.  

Moreover, it is important to acknowledge that the operational efficacy of a well-trained neural network $f: \mathbf{x} \mapsto \mathbf{y}$ is restricted to a very small portion of the vast input space. This limited portion, when compared to the total input space, closely relates to an unknown manifold $\mathcal M$. This manifold can be thought of as emerging from a nonlinear regression process applied to the training dataset $\{\mathbf{x}^{(1)}, \cdots, \mathbf{x}^{(N)}\}$. As a result, there exist infinitely many neural-network functions, denoted as$f_1, f_2 \cdots,$ that demonstrate equivalent efficacy when applied to input data situated near the mysterious manifold $\mathcal M$. This equates to $f_j|_{\mathcal M}\approx f_1|_{\mathcal M}$ for all $j=2,3,\cdots$, despite significant differences between $f_j$ and $f_1$. Moreover, it's noteworthy that within the backpropagation optimization process in a given network, there could be infinitely many global minimizers available.  This observation highlights that, notwithstanding the differences in structural configurations and parameter settings, diverse deep learning networks demonstrate similarily strong performance on input data situated near the training manifold.

Given the previously mentioned constraints, the mapping $f: \mathbf{x} \mapsto \mathbf{y}$ is prone to generating inaccurate outcomes when the input strays even marginally from the unknown data manifold. In this context, the unknown data manifold can be viewed as a mean for data representation and significant dimensionality reduction, expressing the hidden structures inherent in the data. There have been numerous studies on adversarial classification scenarios, such as the erroneous detection of cancer, demonstrating the susceptibility of deep neural networks (developed through a gradient descent-based error reduction process) to assorted disturbances akin to noise. These perturbations can lead to incorrect outputs, a matter of significant concern, especially in critical applications like medical diagnostics. The inherent limitations of deep learning, especially its vulnerability to anomalous inputs that deviate from the training data, are expected to remain difficult to overcome for a considerable period of time. In summary, , while employing the network function restricted to the unknown manifold for significant dimensionality reduction helps to bypass the Curse of Dimensionality, its vulnerability to adversarial attacks presents a fundamentally difficult problem to address.

While unsupervised and self-supervised learning methods have shown promise in the domain of medical image analysis, achieving the level of reliability required for clinical application presents significant challenges. These methodologies, despite their potential, may predominantly remain within the realm of academic research unless substantial advancements are made. 

DL techniques possess the potential to assist doctors in minimizing minor errors, enhancing the accuracy and efficiency of medical diagnostics. However, DL faces inherent challenges in exhibiting the kind of intuitive creativity required for the identification and interpretation of anomalies or novel patterns that have not previously been encountered. Unlike human experts, who can draw upon a wealth of knowledge, experience, and intuition to recognize and explore new phenomena, DL models are constrained by the scope and variety of the data on which they have been trained. These constraints underscore the indispensable need for ongoing human oversight and the integration of DL technologies within a framework that respects and relies on the unique insights and judgment of medical professionals. Such an approach ensures that DL serves as a supportive tool, augmenting rather than attempting to replace the critical decision-making processes in healthcare.

I’d like to conclude this blog with the following thoughts: AI, in my opinion, cannot replace doctors. The expertise of a skilled specialist far surpasses the generalized knowledge of a typical practitioner, and the countless patient-specific cases involve complexities that deep learning models cannot fully capture from limited training data. Similarly, this is why developing self-driving cars that can seamlessly adapt to Manhattan's unpredictable mix of vehicles, pedestrians, cyclists, and pets remains an enormous challenge.