Artificial intelligence (AI) applications in kidney cancer, although still in the early stages, have shown promise in a variety of clinical applications, noted the authors of a review on the topic in the ASCO Educational Book.
“The diagnosis, characterization, management, and treatment of kidney cancer is ripe with areas for improvement that may be met with the promises of artificial intelligence,” wrote Ivan Pedrosa, MD, PhD, of the University of Texas Southwestern Medical Center in Dallas. and co-authors. “Here, we explore the impact of current research work in artificial intelligence for clinicians caring for patients with renal cancer, with a focus on the perspectives of radiologists, pathologists, and urologists.”
In the following interview, Pedrosa, professor of Radiology and vice chair of Radiology Research, discussed some of the promising applications as well as the challenges that need to be overcome.
What is one of the most urgent unmet clinical needs in kidney cancer that artificial intelligence could help address?
Pedrosa: I would emphasize the management of early-stage kidney cancer. Most kidney cancers are diagnosed today when a small renal mass is found incidentally on an imaging study that was done for an unrelated medical condition, such as abdominal pain, trauma, etc. These masses are frequently benign — one in five — and when malignant, may represent one of many different types of cancer. Currently, the management of these small renal masses is challenging because we do not have a robust way to tell what they represent or how they will behave, even with a tissue biopsy.
Ideally, we would want to avoid unnecessary treatment of benign and indolent malignant tumors — the latter particularly in patients with competing comorbidities and limited life expectancy in whom active surveillance may be an option. However, we also want to expedite treatment for aggressive malignant tumors that may put the patient’s life at risk.
Early evidence indicates that AI has the potential to play a role in assisting both physicians and patients to make these decisions. Specifically, AI may improve the characterization of these masses on imaging studies and offer prognostic information that is not well captured today in analyzes of renal biopsies.
In what areas of diagnosing and managing kidney cancer have AI algorithms performed well?
Pedrosa: Artificial intelligence is at its infancy with regard to applications in kidney cancer. Some promising results have been reported in the prediction of histological subtypes based on analysis of imaging studies such as computed tomography and magnetic resonance imaging with deep learning algorithms.
There are also exciting data using AI algorithms to predict molecular alterations such as mutations using digitized histological images. The performance of some of these algorithms is remarkable and, if validated, could have a substantial impact in clinical trials and ultimately, in clinical practice.
In what areas have they not performed as well?
Pedrosa: AI algorithms do well in repetitive tasks for which they are trained. Unfortunately, it is difficult to train algorithms to address the enormous variability we see in the clinic. For example, while AI algorithms perform extremely well for automatic kidney segmentation because the shape of the kidney is more or less predictable, their performance drops for segmentation of actual kidney tumors, which are more variable in size, shape, and appearance. However, we will definitely continue to see progress in this area.
What hurdles need to be overcome for AI to realize its full potential?
Pedrosa: AI algorithms need adequately curated and annotated data to be trained. Their performance suffers when the training data is inconsistent or the algorithm encounters scenarios that were not included in the training dataset. Unfortunately, there are innumerable sources of variability in kidney cancer patients, and training of AI algorithms for those would require enormous curated datasets. Examples of these include differences in patient body habitus, imaging technique, imaging acquisition protocols, histologic subtype, tissue processing, molecular clusters, etc.
The use of AI algorithms in clinical practice will likely expand when we figure out a way to generate large well-curated datasets that are diverse enough so that AI algorithms can be trained to recognize most clinical scenarios. In the meantime, AI algorithms may provide workflow improvements to ease some of the repetitive, time-consuming tasks done by humans such as image annotation.
What do you envision the role of AI in diagnosing and managing kidney cancer will be in 10 or 20 years?
Pedrosa: I believe AI will be at the center of deciphering tumor heterogeneity — arguably the most difficult challenge in kidney cancer, and perhaps oncology in general, today. A number of new and exciting therapeutic regimens have been approved in recent years for patients with locally advanced and metastatic disease. Thankfully, we will continue to see more options becoming available.
However, currently the choice between these regimens is largely empirical. Although emerging data suggest that some molecular signatures may indicate a higher likelihood of response to some of these regimens, this determination requires tissue samples.
Unfortunately, we know that kidney cancer is highly heterogeneous and it is quite possible that a single biopsy is not representative of the entire tumor burden, at least in some patients. Although we cannot obtain tissue from every metastatic site, other approaches such as imaging or liquid biopsies have the potential to evaluate the entire tumor burden. This is extremely enabling even today, when we combine systemic therapy with local interventions such as surgery, ablation, or radiation in patients experiencing heterogeneous responses.
However, the amount of information that we need to process to make these decisions and to evaluate the oncologic outcomes at the patient and population level is overwhelming. I believe artificial intelligence will play a critical role in integrating all these data — imaging, pathology, laboratory, molecular, omics, etc. — to develop novel multidimensional predictive and prognostic biomarkers that assist in the management of kidney cancer patients, from their early diagnosis to the metastatic disease.
Read the study here and expert commentary about the clinical implications here.