Andrew Janowczyk
Assistant Professor, Department of Biomedical Engineering, Emory University and Georgia Institute of Technology, Atlanta, Georgia, USA; Department of Oncology, Division of Precision Oncology, Department of Diagnostics, Division of Clinical Pathology, Geneva University Hospitals, Switzerland
Controversial opinion? One of my most controversial opinions about the field of pathology, though perhaps not so controversial among insiders, concerns the media's portrayal of machine and deep learning research, or more generally “Artificial Intelligence”. The way these technologies are presented often creates a discordant narrative compared to the realities we face daily as practitioners.
In many media reports, machine and deep learning technologies in pathology are described as being "intelligent," imbued with human-like characteristics. However, I believe that these technologies have not yet achieved that level of sophistication. A humorous illustration of this can be found in the hypothetical scenario of a deep learning classifier tasked with identifying whether a digital tissue image contains cancer. While the classifier might perform admirably with actual tissue images, what if a picture of a cat is accidentally submitted? The classifier would still produce a binary yes/no response, and not “think” anything strange of it. However, a genuinely intelligent entity would recognize something is amiss. At the very least, it would raise a red flag, if not reject the premise outright.
This example underscores a critical point: while the potential impact of these technologies is undoubtedly significant, we must be careful in how we present them to those outside the field. Overstating their capabilities can lead to misconceptions and an underestimation of the inherent biases and challenges involved. These technologies are powerful tools, but they are not infallible and still require careful oversight and contextual understanding.
Thus, it's crucial for us as practitioners to communicate the realities of these advancements accurately. By doing so, we can foster a more informed and realistic perspective among non-practitioners, mitigating the risk of naivety and ensuring a balanced view of both the capabilities and limitations of machine and deep learning in pathology.
Making the most of an invention or innovation? The secret to maximizing your invention's potential is to share and discuss it with a diverse range of colleagues and professionals. Often, we underestimate our invention's impact because we view it only within our specific context. By presenting your idea in different venues and to individuals from various backgrounds, you gain fresh perspectives and uncover new opportunities that you might not have considered. This approach broadens the scope of your invention's impact, fosters interdisciplinary innovation, and increases the chances of successful adoption across various domains.
Deciding which problems to tackle?My approach is heavily guided by discussions with pathologists and clinicians, who are the end users of any solutions we develop. Their insights are crucial in ensuring that the problems I focus on are genuine, real-world issues rather than hypothetical scenarios I might create in isolation. By engaging with these professionals, I can better understand their needs and the practical challenges they face, allowing me to direct my efforts towards solving meaningful and impactful problems.
Additionally, more and more I am prioritizing “pre-experimental” rigorous scientific methodology in my work. After formulating a hypothesis, I delve deeply into the available data to ascertain whether the hypothesis is justified and if the problem at hand is actually answerable with the data we have. This involves thoroughly checking for potential issues such as batch effects, which can significantly skew results and diminish the impact of our work.
By investing time in this preliminary analysis, I can filter out many problems that, despite initial appearances and great interest/enthusiasm to tackle, are not solvable given the constraints. This careful upfront work helps avoid the common pitfall of reaching the end of an experiment only to realize that critical concepts were overlooked. In essence, this approach not only streamlines the research process but also ensures that the solutions developed are robust, reliable, and directly address the needs of the end users, thereby maximizing the real-world applicability and impact of the work.
