November 6, 2025 | For most physicians, concerns about artificial intelligence (AI) continue to outweigh enthusiasm for its potential to improve healthcare decision-making and patient outcomes. Doctors need to understand the reasons behind the predictions of AI tools if they’re expected to use the insights as a starting point for discussing treatment options with their patients, says Gustavo Carneiro, Ph.D., professor of AI and Machine Learning at the University of Surrey’s Centre for Vision, Speech and Signal Processing (CVSSP).
He and his colleagues have stepped into the gap with a predictive algorithm for osteoarthritis (OA) progression that offers interpretability by showing which anatomical knee landmarks on X-ray images the AI is monitoring for telltale changes. The end game is an interactive tool that can be used like ChatGPT for early detection and monitoring of a variety of conditions, including breast and lung cancers, using medical images, Carneiro says.
Many other deep learning tools are under development offering a numerical score tied to the likelihood of disease progression, according to David Butler, a postgraduate research student at Surrey’s CVSSP. But this may well be the first interpretable machine learning method that pairs predictive scores, presented as a percentage, with annotated “future images” for risk estimation.
A generative diffusion model was used in the latest study to predict knee OA severity 12 months out and offer 16 key points of interest on generated images as a visual representation of how the situation may have evolved, Butler explains. The modeling technique requires far fewer artificial neurons than would be needed for more complex classification exercises or the production of natural images with more feature diversity, which translates into faster image generation as well as outputs that are easier to understand.
When applied to the Osteoarthritis Initiative dataset, the approach outperformed state-of-the-art AI tools for predicting OA progression by a modest 2% but did so around nine times faster as well as improved the interpretability by generating both future images and localizing anatomical knee landmarks (MICCAI 2025, DOI: 10.1007/978-3-032-05185-1_52). The study was presented at the recent International Conference on Medical Image Computing and Computer Assisted Intervention in Daejeon, South Korea.
Industry sponsors have expressed interest in using the novel AI tool in upcoming OA interventional clinical trials, Butler reports. This is because the current method for evaluating drugs for regulatory approval involve a six-minute walking test that is time consuming and can be difficult for people with osteoarthritis due to pain, exhaustion, and functional impairment.
Creation of the predictive model was part of Butler’s Ph.D. work, Carneiro points out, and the pair already have another paper underway exploring its potential applications beyond OA. Intellectual property discussions have also begun with the university regarding how the AI method might be commercialized.
The comparator AI systems used in the latest study were classification models, most of which simply input images and output risk scores, says Butler. Several tools also produce images using NVIDIA’s StyleGAN but aren’t concerned with predicting disease progression.
The novelty of the AI tool is that it provides clinicians with meaningful information they can use to frame conversations with OA patients who need treatment and can’t readily appreciate the consequences of doing nothing, says Carneiro. Doctors can more easily trust the risk assessment because it comes with the explainability of now-and-later X-ray images. “The whole point of this tool is to involve both the doctor and the patient in the process of accepting the diagnosis ... and the treatment that the doctor wants to prescribe.”
The risk of disease progression is forecast out one year because that’s the time horizon between images captured in the Osteoarthritis Initiative dataset, which is a public research database containing longitudinal data on nearly 5,000 participants from a multi-center, observational study on OA. The next step is to start working with doctors to confirm that 12 months is the ideal benchmark for disease progression predictions, Carneiro says.
Prediction tools are too new for a “golden rule” on the matter, adds Butler. If these tools were better at longer time scales, without creating uncertainty about the outputs, it might make sense to extend them.
Ultimately, AI tools need to integrate with the workflow of physicians, which is also “a little bit of a complication,” says Carneiro. “As long as they see the value in a tool, they may be more willing to adopt it.”
With the product they have under development, markers on the future images relate to radiographic features used by the Kellgren-Lawrence scale to assign OA severity a grade from 0 (none) to 4 (severe) based on the presence, size, and severity of these characteristics, Butler says. These are characteristics labeled by radiologists on images in the dataset, and the basis of the model’s prediction scores and landmarks.
The main inputs are joint space narrowing indicative of cartilage loss and the presence of bony overgrowths, he continues. From an original image, the system predicts what the image is going to look like in a year in parallel with the 16 markers and the risk score. Everything is correlated because the algorithm was trained on outcomes at that future time point.
“If the landmarks don’t make sense and map up with the image, then there’s an issue,” says Butler. “We also found that predicting the landmarks increases the accuracy of the risk prediction.”
Osteoarthritis is only the beginning, says Carneiro, and the starting point because of the dataset that was on hand. The development team also intends to use the technique on mammograms.
Past work on the prediction of lung cancer began with an imaging dataset on patients without any symptoms or indication of having the condition, putting the predictive focus on the “bigger outcome” of cancer development that encompasses the entire disease life cycle, he adds. The modeling technique is “definitely not constrained to osteoarthritis; it should go well beyond that.”
The long-term goal is to have an interactive AI tool that will answer questions doctors ask using images of what a person’s condition will look like at a future time point—and how that might change depending on the treatment that gets administered, Carneiro says. Talks are now underway with other universities, and forthcoming ones will be with potential industry partners to aid in the tool’s market translation, he adds.
Plans are to submit the latest study for publication to one of the main engineering journals by the end of the year, says Carneiro, who expects 2026 will bring several research milestones. For any collaborations with doctors on more applied AI projects, publication ambitions would shift to medicine and biomedical research journals.
“Integrating AI into the workflow of clinicians and making it more trustworthy for patients is very important,” Carneiro stresses. “AI has focused way too much on the data and not enough on the people. We need to change our focus to who is going to use the tool and who is going to be affected by [it], and our tool is one example of that.”