(UroToday.com) Dr. Peter Chang continued the morning plenary session with an excellent presentation on Artificial Intelligence (AI). He began by first defining AI as either traditional machine learning whereby logic had to be programmed in by a human expert versus newer computational deep learning where the AI is provided with an end goal that it then tries to achieve through standardized algorithms. Dr. Chang provided the IBM Deep Blue AI as an example of traditional AI in which the AI algorithms were trained by hundreds of human chess experts to play chess. The IBM deep blue AI played chess against the then world champion and most accomplished chess player Gary Kasparov in 1997. It was thought that Kasparov could anticipate 10 moves in advance and in contrast the IBM Blue AI could anticipate hundreds and thousands of moves every second, ultimately beating the reigning human chess champion.
Dr. Chang then contrasted with an explanation of the newer deep learning AI’s where the algorithms started not with logic or instructions but with a final result or end task that was to be achieved. He explained that from there on the training process was very hands offs. He provided an example of the game ‘Go’ which has more moves than atoms in the entire universe. For the longest time, it was thought that for a complex game like Go, there would be a requirement for a human expert to master the game however with the onset of deep learning, an AI algorithm was developed which was pitted against itself which resulted in the AI AlphaGo which beat the then reigning Go champion Lee Sedol from South Korea in 2016. The AIphaGo algorithm played 3 million games against itself in just 3 days developing strategies that even human experts could not comprehend. When pitted against a human it appeared the AI was losing with the first half of moves on purpose before securing the win in the end. Deep learning AI is the newer flavor of machine learning that has been dominant in the past 10 years. Dr. Chang explains that deep learning does not require any explicit instructions or logic rather contains a network of neurons and instruction sets that code for how weak or strong an interaction may be.
With this, he moved on to what is possible with neural networks in medicine. He highlighted the three major domains in medicine where AI neural networks may have clinical applications. Firstly, was the detection of abnormalities such as identifying a stone or tumor on cross sectional imaging, the second was extracting information from the area of interest such as identifying the volume of a stone or tumor or even calculating something like a BIRADS score from an ultrasound image and lastly was information around prognostication i.e which therapy might result in the best outcomes based on the information AI has been able to capture. Dr. Chang then outlined current examples of these AI in action. He presented images from an AI algorithm designed at the University of California, Irvine which could examine CT scans and identify renal masses from surrounding normal renal tissue. Not only could the AI clearly identify the borders of the tumors but also identify further granular information such as the volume of the tumor and proximity to important structures.
He then presented data from an AI algorithm that could read prostate MRIs and segment prostate lesions and then assign a PIRADS score. He went one step further to show clinical applications of this AI such as differentiating between low risk and high-risk prostate cancer and explained that if we had enough data to feed the deep learning algorithms outcomes such as the above could also be achieved.
Dr. Chang then moved on to his final example of AI based augmented reality models of renal tumors that were currently being used by the University of California, Irvine in which renal tumors were segmented by AI and allowed the surgeon to in real time examine the tumors and study their characteristics and association with surrounding anatomical structures prior to operative management.
He then concluded with statements about the future of AI in medicine. He explained that despite all the above innovations in AI, AI is generally underutilized in clinical medicine and that current algorithms in use in certain hospitals are not readily transportable to other hospitals. He also noted that some AI algorithms that may produce a great amount of academic data may not necessarily have clinical implications. For example, if AI can identify a benign renal lesion on cross sectional imaging with a 90% certainty would that change a clinician's decision making process given the 10% risk of missing a malignancy? Dr. Chang concluded that he did not think AI can replace humans at this point in time but inf the future may help clinicians by automating certain tasks and improving clinical workflows.
Presented by: Peter Chang, MD, University of California, IrvineWritten by: Sohrab Naushad Ali, MD, MSc, FRCSC, Assistant Clinical Professor, Department of Urology, University of California Irvine, @sohrabnaushad on Twitter during the 39th World Congress of Endo urology and Uro-Technology (WCET), Oct 1 - 4, 2022, San Diego, California.