“It’s hard to see things when you’re too close. Take a step back and look.” - Bob Ross
(UroToday.com) It is nearly impossible to fathom a world devoid of the Internet and cell phones. Yet, the full extent of one innovation’s impact often dawns on us only after its full integration into our daily lives. Initially, these advancements may appear raw and unpolished, but the path of creativity has a thousand steps. In retrospect, the era spanning from 2023-2024 will be remembered as the dawn of the Artificial Intelligence Revolution, echoing the Internet’s emergence in the 1960s.
In just a year, multiple AI algorithms have become widely available to the public. From the meteoric rise of ChatGPT to competitors like Bing AI, Google AI, DALL-E, and SORA, vying for prominence, the landscape is burgeoning with innovation. Yet, as with any new field, the lexicon of artificial intelligence often proves overwhelming and more often than not, confusing.
- Machine Learning vs. Deep Learning
Machine Learning is a subtype of AI through which algorithms can analyze large datasets, identify subtle patterns within the dataset, and learn from those patterns to make predictions. Deep learning (DL), on the other hand, is usually associated with a more complex architecture, with AI units interconnected like neurons in a multi-layer functional system (hence “deep” learning). When discussing DL models, one should be familiar with the enigma of the “black box problem”, wherein the decision-making process is opaque, meaning that we cannot see how DL systems make their decisions.
- Supervised vs. Unsupervised AI
Machine Learning can be “supervised”, in which the input dataset must be manually labeled (i.e., mpMRI of the prostate pre-processed to teach the algorithm how cancer looks like on imaging) or “unsupervised” in which the dataset is unlabeled (i.e., multiple mpMRI of the prostates are fed to the AI algorithm with the program teaching itself the rules of the game).
- Transformers and Large language models:
Large language models (LLMs) are a sub-type of deep learning AI specialized in comprehending natural text written by humans and generating human-like responses. Powered by transformers, these algorithms analyze vast amounts of text data (in the case of ChatGPT the entire content of the Internet available in English) to emulate how a human would respond to a question. The most well-known large language model to date is ChatGPT.
Figure 1. Artificial intelligence. Types and Subtypes. (Source: Nadia BECHANE (M2 IESCI, 2018).
Dr. Geoffrey Sonn (Stanford Urology) subsequently explained AI’s potential ramifications in medical imaging review, particularly in the realm of multiparametric MRI. While AI augments image acquisition speed and interpretation efficiency, its generalizability and accuracy largely rely on the training dataset and its segregation from testing and validation datasets.
While for new medications and intervention validation randomized clinical trials have the strongest evidence strength, the gold standard for AI algorithm generalizability is international competitions or “grand challenges (such as the Prostate X Challenge or the PI-CAI Challenge)”. In these challenges, different algorithms are compared against each other in a bias-free manner, using the same training and testing data. These endeavors foster unbiased comparisons between diverse AI solutions, addressing concerns regarding trust and comparability.
Figure 2. A novel model for AI model external validation-creating “grand challenges”, where different AI algorithms designed for the same purpose are tested against the same testing dataset.
Furthermore, the accuracy of AI-based imaging interpretations is contingent upon precise labeling, necessitating extensive expert input. But getting experts to fully train an AI algorithm, often with hundreds or if not thousands of labels might not be easily feasible, especially since experts represent a limited group and time is a finite resource. Given the logistical hurdles associated with this endeavor, multi-institutional collaborations emerge as indispensable.
However, as Dr. Joseph Liao (Stanford Urology) highlighted; AI's implications are not limited to medical imaging interpretation. Through its neural network-like architecture, AI discerns intricate patterns and correlations across disparate datasets, amalgamating longitudinal histories with biomarkers, imaging, and pathology results.
Lastly, Dr. Prokar Dasgupta (King’s College of London Urology), concluded this fascinating plenary session by highlighting the evolving litmus test for AI’s societal impact. While in the 1960s, Alan Turin’s query of “Can Machines Think?” defined an AI algorithm’s worth, contemporary scrutiny pivots towards the Weizenbaum test “How does AI impact societies?”.
Democratizing artificial intelligence is essential. Dr. Dasgupta’s proposition of the 3Cs-Countries, Companies, and Civil society underscores the collaborative nature of this endeavor. Only through global collaboration between academic, industrial, and societal stakeholders can AI truly catalyze transformative change while navigating the ethical nuances encapsulated in the Weizenbaum test: How does AI influence our quotidian existence?
Moderator: Andrew Hung, MD, Cedars Sinai Urology, Los Angeles, CA, USA
Panelists: Geoffrey Sonn, MD, Stanford Urology, CA, USA; Joseph Liao, MD, Stanford Urology, CA, USA; Prokar Dasgupta, OBE, FRCS, King’s College of London, London, UK
Written by: Andrei D. Cumpanas, Incoming PGY-1 Urology Resident, Department of Urology, University of California Irvine, @andreicumpanas on X during the 2024 American Urological Association (AUA) Annual Meeting, San Antonio, TX, Fri, May 3 – Mon, May 6, 2024.