Deep learning (DL) dose prediction models aim to predict the radiation dose that a patient will receive before the treatment planning process even begins. By taking only the patient's anatomy (organ contours) as input, these models can instantly produce an estimated dose distribution -- giving clinicians a head start on setting treatment goals.
However, a fundamental challenge limits these models from being widely deployed: planning style variation. In radiation oncology, different physicians, dosimetrists, and institutions make different trade-off decisions -- such as how aggressively to spare the bladder versus the rectum -- even when following the same clinical guidelines. As a result, a model trained at one institution or on one planning style may produce systematically wrong predictions when applied elsewhere.
This problem is particularly acute because building a new deep learning model from scratch requires collecting and curating hundreds of high-quality patient cases -- an enormous effort that most institutions cannot undertake independently. A practical solution would allow a pre-built model to be quickly adapted to a new setting using only a small number of local cases.
Researchers from UT Southwestern Medical Center investigated whether transfer learning -- a technique in which a pre-trained AI model is fine-tuned on a small new dataset -- could solve the planning style generalizability problem for prostate cancer dose prediction.
The team built a single source model trained on 108 patients from one institution using one predominant planning style (the "conformal" style). They then tested whether this model could be adapted to three different planning styles used at the same institution (Internal-A, B, and C) and one planning style used at an entirely different institution (External) -- each using only 14 to 29 new training cases.
This approach is clinically important because it means an institution could receive a pre-built model from a larger center and adapt it to their own local practice with a modest local dataset, rather than needing to build a model entirely from scratch.
The dose prediction model was built on a 3D UNet architecture, a deep convolutional neural network widely used in medical imaging. Its input consisted of binary masks (structural outlines) for the planning target volume (prostate tumor), bladder, rectum, femoral heads, and body. The model's output was a predicted 3D radiation dose distribution across the entire treatment volume.
The four target planning styles represented genuinely different clinical trade-off strategies. Internal-A emphasized lower femur head doses; Internal-B prioritized rectal sparing over bladder and femur sparing; Internal-C aggressively spared both bladder and rectum at the cost of higher femur doses. The External style used a completely different treatment planning system (Erasmus-iCycle) and different beam configuration, producing distinctly different low-dose patterns across the patient body.
For transfer learning, the approach was elegant in its simplicity: the first half of the UNet (which extracts anatomical features) was frozen (kept unchanged), while only the second half (which maps features to dose distributions) was retrained on the new planning style's data. The final network layer was randomly re-initialized to help the network escape local optima. The learning rate was reduced to one-tenth of the original to allow careful, stable fine-tuning.
The source model performed well on its own planning style but produced significantly inaccurate predictions when applied to other styles. For the External institution style -- which was most different from the source -- the source model achieved a gamma passing rate of only 0.47, while the adapted target model reached 0.84, a dramatic improvement (p less than 0.01).
Spatial agreement, measured by Dice Similarity Coefficient (DSC) across isodose volumes, improved consistently with transfer learning. For the External style, the target model achieved up to a 10% DSC improvement in low-dose regions and 5-8% improvements in high and intermediate dose volumes -- all statistically significant.
For key dose-volume histogram (DVH) parameters in the External style, transfer learning reduced prediction errors from as high as 13.5% (source model) to within 1.6% across all tumor and organ metrics. This level of accuracy is clinically meaningful and approaches what would be expected from a model specifically trained on that institution's data from scratch.
The authors identified several practical clinical applications for adapted dose prediction models. First, the predicted dose can serve as a physician planning directive tool -- at UT Southwestern, physicians trigger the prediction at the time of contour completion and use it to review planned objectives and make adjustments before submitting instructions to the dosimetrist, saving iterative back-and-forth communication.
Second, the tool can support treatment modality decisions (for example, comparing expected dose distributions for 3D conformal radiation versus IMRT), enabling more informed clinical decisions earlier in the planning process. Third, predicted doses can facilitate efficient peer review quality checks, allowing a second reviewer to quickly assess whether a patient's treatment plan is within expected norms.
The method produces what the authors describe as a "physician-acceptable" plan rather than a theoretically optimal one -- reflecting the reality that training data comes from clinically approved cases. This practical framing makes the tool more relevant and interpretable to clinicians who must integrate it into their existing workflows.
The study highlights a core reality in clinical oncology: even when institutions follow the same national guidelines, the individual preferences of physicians and dosimetrists produce measurably different dose distributions. A DL model that ignores this heterogeneity is likely to produce predictions that are confusing or unhelpful to local clinicians.
Transfer learning offers an elegant solution to what the authors identify as a key barrier to widespread AI adoption in radiation oncology: the need for large, curated, institution-specific training datasets. By requiring only 14-29 adapted cases, the method is realistic for nearly any clinical center to implement, removing a major practical obstacle.
Future directions include developing Bayesian uncertainty estimation -- allowing the model to not only predict a dose distribution but also to express its confidence in the prediction. This would allow clinicians to see which parts of the predicted dose are reliable and which may require additional scrutiny, making the AI tool more transparent and safer for routine clinical use.