“Inside Access” provides interviews and background on open access articles from the Journal of Interventional Radiology.
Kuhn TN, et al. Artificial intelligence–driven patient selection for preoperative portal vein embolization for patients with colorectal cancer liver metastases. J Vasc Interv Radiol. 2025 Mar;36(3):477-488. doi: 0.1016/j.jvir.2024.11.025.
Read the full article on JVIR.
Tell us about you, your team and your institution.
Tom N. Kuhn, MD: I'm currently a clinician scientist at the German Cancer Research Center with a focus on machine-learning-based predictions on clinical decision-making and procedure outcomes. Our team consists of interdisciplinary experts in the field of interventional radiology, oncology and biomedical engineering. We're working across multiple high-volume cancer centers. Our institution aims to integrate machine learning into clinical workflow to optimize patient selection and therefore improve patient outcomes.
Why did you choose to pursue this specific topic?
TK: We thought that portal vein embolization is a very important treatment, which is crucial for patients with colorectal liver metastases who require a liver resection but have insufficient future liver remnant (FLR%). However, it is currently difficult to select the right candidates due to variability in patients’ response. Traditional methods were very limited and based on clinical parameters. We tried, by integrating additional features into machine learning algorithms, to improve patient selection accuracy and therefore optimize surgical outcomes.
What key factors determine whether a patient with colorectal cancer liver metastases would be eligible for a successful PVE treatment?
TK: That very much depends on the patient. For our cohort, most of our patients were heavily pretreated with chemotherapy, meaning they received chemotherapy for more than 12 weeks. Therefore, it was defined that the FLR% should be above 30% based on the paper by Shindoh et al.1 To identify key factors, we performed a feature importance analysis of our machine learning algorithms. For our patients, it was very important to look back on what features were important to make accurate predictions. If we look at the top 10 features that made our model accurate for the FLR% model, there were two markers that you generally use in clinical practice. For example, the baseline FLR% or the baseline left-liver volume. But the rest were features that are currently not used for clinical decision-making, meaning there were six radiomic features that were extracted from the CT scans and two features from our statistical shape model approach where we used principal components as features. This underlies the importance of machine learning in this specific setting.
How did the machine learning model help identify the patients that were chosen?
TK: That was very much based on the specific features and how important they were—meaning mostly features that are not visible to the naked eye were predictive for their specific model. And that's basically the importance of machine learning in that specific setting.
How did machine learning and radiomics improve predictions compared to traditional methods that only considered a small number of clinical and imaging factors?
TK: What we have seen before is that for that specific setting, radiomic features were very important for prediction accuracy and what we could show as well. What I think is a novelty of our study is that a statistical shape model, which helps to quantify liver morphology and their growth dynamics, helped in making those predictions as well. And those are all features that are usually not used before, and I think that really underlies the importance of machine learning to analyze complex patterns, especially across multiple variables.
What are the advantages of incorporating detailed imaging features like liver shape in the predictive model?
TK: The importance of it lies in its predictive value, meaning that those features are very important for our machine-learning algorithms. And that was true not only for the FLR% model but also the kinetic growth percentage model and the total liver volume as well.
The study mentions that this model was tested at multiple institutions. How did it perform across these different hospitals?
TK: Our model was trained from four high-volume cancer centers, three from the U.S. and one from Japan. We used three institutions for training and testing and then we kept one institution separate for external validation. We were able to show that external validation did not reveal any statistically significant difference, meaning that the external testing was basically as good as the internal testing. This indicated the generalizability of our model. I think that is very important, because most of the studies in general do not have an external validation, and this is basically a real-world test from a separate institution that helps to underline the performance of a model under real-world conditions.
How reliable is it for predicting treatment outcomes in patients from diverse backgrounds or with varying medical histories?
TK: Based on our data, it is reliable. So, for example, one of our institutions was from Japan and they perform PVE very differently than it is performed in the U.S. Under these circumstances, there was no real difference between the institutions on key markers or accuracy. Therefore, our model was very robust based on the diverse backgrounds of patients.
Do you and your team have any next steps or plans for follow-up with this study?
TK We definitely do, though I think the importance of this study lies in its applicability. One thing that we are working currently on is integrating that into real-world clinical decisions. So basically, clinicians can use that tool to make predictions. I think what is important, too, is that the basis of this machine learning algorithm can be used and applied to different other techniques of liver intervention, such as ALPPS or radiation lobectomy. It could be used for HCC patients as well.
Is there anything else you feel readers should know?
TK: What makes our work truly remarkable is its practical applicability. For instance, take our AI-driven liver model—it enables predictive simulations, offering a baseline liver shape alongside projections for 2 and 4 weeks after a hypothetical portal vein embolization (PVE). This means that without performing any procedures, we can visualize how the liver might change over time, providing surgeons with invaluable insights for preoperative planning.
Moreover, our model allows for a dynamic 4D simulation, illustrating the liver’s morphological evolution in real time. This capability is particularly exciting, as it grants clinicians a powerful tool to assess potential outcomes before any intervention takes place—ultimately aiding in the selection of patients who would benefit most from PVE.
References
- Shindoh J et al. Optimal future liver remnant in patients treated with extensive preoperative chemotherapy for colorectal liver metastases. Ann Surg Oncol. 2013 Aug;20(8):2493–2500. doi: 10.1245/s10434-012-2864-7. ePub 2013 Feb 3. PMID: 23377564; PMCID: PMC3855465.
