Use of AI after kidney transplantation

Jesper Kers is a renal pathologist and researcher at Amsterdam UMC and Leiden UMC. Together with his research group, he focuses on developing and training computer models. Here, he talks about the use of AI in his field.

Jesper Kers: 'When we think about rejection or other problems after a kidney transplant, a kidney biopsy is often taken. In my research group, we are looking at whether pathologists can be supported by computer models when assessing those biopsies.'

Training model with 5,800 scans

'We have developed so-called convolutional neural networks, complex models consisting of millions of mathematical formulas. Using 5,800 scans of kidney biopsies, we trained such a model, and it turns out to be quite capable of distinguishing between normal findings and signs of rejection.'

Improving diagnostics

'Currently, there is sometimes a difference of opinion between experts about what a biopsy shows. So there is room for improvement. With the help of AI, we may be able to improve diagnostics, which could ultimately lead to more effective treatments.'

Research needed: is a pathologist with AI better?

'Before we can apply AI clinically, we must conduct randomized studies. Those studies can show whether a pathologist with AI support is more accurate and consistent than without. If that turns out to be the case, I think AI could be used in daily practice within 5 years.'

New models

'We are also developing new models. An important goal is to integrate different types of data: clinical, molecular, and histological data. That could really become a second pair of eyes for the doctor.'

Challenges and questions

'Developments in this field are moving fast. That is great, but it also brings challenges. Do we understand enough about how a model reaches its conclusions? The algorithms convert pixel groups into a kind of barcode, but we don't know exactly how that translation works. Some people get nervous about that.'

'At the same time, we are extensively validating the models with data from multiple hospitals. If those tests show that a model is reliable, does it still matter that we don't understand all the details? That remains an interesting question in this field.'