📄 medicine

Dosiomics-based prediction of late genitourinary toxicity after Cyberknife for prostate cancer: a SHAP-explainable machine learning approach

This study demonstrates that an explainable machine learning model utilizing organ-specific dosiomics features from CyberKnife SBRT plans can effectively predict late genitourinary toxicity in prostate cancer patients, achieving high accuracy while identifying clinically interpretable risk drivers through SHAP analysis.

Original authors: Rocco Mottareale, Marcello Serra, Donato Pezzulla, Savino Cilla, Valentina Borzillo, Esmeralda Scipilliti, Dario Franzese, Sisto Perdonà, Francesco Passaro, Alessandro Izzo, Sabrina Rossetti, Sandro P

Published 2026-07-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rocco Mottareale, Marcello Serra, Donato Pezzulla, Savino Cilla, Valentina Borzillo, Esmeralda Scipilliti, Dario Franzese, Sisto Perdonà, Francesco Passaro, Alessandro Izzo, Sabrina Rossetti, Sandro Pignata, Gianluca Ametrano, Valentina d'Alesio, Francesca Buonanno, Cecilia Arrichiello, Simona Mercogliano, Rossella Di Franco, Vincenzo Ravo

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the "Aftermath" of Treatment

Imagine a patient with prostate cancer is about to undergo a very precise, high-tech radiation treatment called CyberKnife. This treatment is like a sniper rifle for cancer: it delivers a massive dose of radiation in just five sessions to kill the tumor while trying to miss the surrounding healthy organs.

While the treatment is excellent at curing the cancer, there is a risk of "collateral damage" later on. Specifically, the patient might develop late urinary problems (like trouble peeing or pain) months after the treatment is finished. Doctors want to know before they start which patients are likely to have these problems so they can adjust the plan.

This paper asks: Can we use a computer to look at the radiation map and predict who will get sick later?

The "Dosiomics" Recipe: Turning Radiation into Data

Usually, doctors look at a radiation plan and check simple numbers (like "how much total radiation hit the bladder?"). This paper tried something more complex called Dosiomics.

Think of the radiation dose distribution not just as a number, but as a 3D landscape or a topographic map.

Standard approach: Checking the average height of the mountains in the map.
Dosiomics approach: Looking at the texture of the mountains. Are they smooth? Are they jagged? Are there sudden cliffs? Are the valleys deep?

The researchers used a computer program (PyRadiomics) to turn the radiation maps of the prostate, bladder, and rectum into thousands of tiny data points describing these "landscapes." They were looking for hidden patterns in the shape and texture of the radiation that simple numbers miss.

The Detective Work: Machine Learning & SHAP

The team fed this mountain of data into a smart computer program (an XGBoost machine learning model). Think of this model as a super-detective that has read millions of radiation maps and learned to spot the subtle clues that lead to future urinary trouble.

However, AI is often a "black box"—it gives an answer, but you don't know why. To fix this, the researchers used a tool called SHAP.

The Analogy: Imagine a group of detectives trying to solve a crime. SHAP is like a scoreboard that shows exactly how much each clue contributed to the final verdict. Did the "jagged texture of the bladder" add 10 points to the risk? Did the "shape of the prostate" subtract 5 points?
This made the computer's decision explainable, allowing doctors to see exactly which features were driving the prediction.

What They Found

The researchers studied 87 patients. About 17% of them developed urinary problems later on.

The Prediction: The computer model was quite good at spotting who would not get sick (it was very accurate at saying "safe"). It was a bit harder for the computer to catch the few people who would get sick (because there were so few of them), but when it did predict a risk, it was usually right.
The Clues (The "Why"): The SHAP tool revealed the specific "clues" the computer was using:
- Bladder Shape: If the bladder was wider or flatter in a specific direction, the risk went up. It's like a balloon that is stretched out; the radiation might hit a sensitive spot more easily.
- Bladder Texture: If the radiation pattern inside the bladder was "patchy" or "fragmented" (like a mosaic with uneven tiles) rather than smooth, the risk increased. This suggests that scattered "hot spots" of radiation are dangerous.
- Prostate Texture: If the radiation inside the prostate was too uniform (too smooth), it actually signaled a higher risk. This suggests that the specific "spiky" or varied nature of CyberKnife radiation is usually a good thing, and losing that variation might mean the plan isn't optimized.

The Conclusion

The study concludes that by treating the radiation dose like a complex image (Dosiomics) and using an explainable AI (SHAP), they can predict late urinary toxicity better than looking at simple numbers alone.

The Catch:
The study was small (only 87 patients) and done at just one hospital. The authors are careful to say this is a "proof of concept." They haven't proven this works for everyone yet, and they need to test it on much larger groups of people in different hospitals before it can be used as a standard tool in clinics.

In short: They built a smart, explainable system that looks at the "texture" of radiation maps to guess who might have urinary trouble later, finding that the shape of the bladder and the patchiness of the radiation dose are key warning signs.

Technical Summary: Dosiomics-based Prediction of Late Genitourinary Toxicity after CyberKnife for Prostate Cancer

Problem Statement
Prostate cancer (PCa) is frequently treated with Stereotactic Body Radiotherapy (SBRT), particularly using the CyberKnife (CK) platform, which offers high conformality and steep dose gradients. While SBRT provides excellent biochemical control, late genitourinary (GU) toxicity remains a significant clinical concern that can impair quality of life. Traditional predictive models often rely on dose-volume histograms (DVHs), which may fail to capture the complex spatial heterogeneity and dose-anatomy interactions inherent in CK plans. There is a need for predictive tools that can identify patients at risk for late toxicity (onset >6 months, grade >1) by leveraging the full spatial information of the dose distribution, while also providing interpretable insights to guide treatment planning.

Methodology
The study employed a retrospective analysis of 87 patients with localized prostate adenocarcinoma treated with CK SBRT (35–36.25 Gy in five fractions). The workflow integrated dosiomics, machine learning, and explainable AI (XAI):

Data Extraction: DICOM RTDOSE and RTSTRUCT files were used to extract dosiomic features from five regions of interest (ROIs): prostate, planning target volume (PTV), bladder, rectum, and penile bulb. Using PyRadiomics, 76 features per ROI were extracted, categorized into shape/geometric descriptors, first-order statistics, and texture features (GLCM and GLRLM), resulting in a total of 380 initial features.
Feature Selection: A rigorous filtering pipeline was applied to the training set (80% of patients). This included:
- Removal of highly collinear features (Pearson correlation $|\rho| > 0.95$ ).
- Exclusion of quasi-constant features (Coefficient of Variation < 0.05).
- LASSO (Least Absolute Shrinkage and Selection Operator) regression to select a sparse subset of predictors.
- This process reduced the feature space by approximately 95%, retaining 20 predictors.
Model Development: An XGBoost (Extreme Gradient Boosting) classifier was trained to predict the binary outcome of late GU toxicity. The model utilized group-aware 5-fold cross-validation (GroupKFold) to ensure patient-level independence. Class imbalance was addressed using the scale_pos_weight parameter rather than synthetic oversampling.
Explainability: SHapley Additive exPlanations (SHAP) were employed to interpret the model. This included global analysis (ranking feature importance via mean absolute SHAP values) and local analysis (decomposing individual patient predictions to identify specific risk drivers).

Key Results

Incidence: Late GU toxicity (grade >1, >6 months) occurred in 17.24% of the cohort.
Performance: On an independent hold-out test set (20% of patients), the model achieved a mean ROC-AUC of 0.85 ± 0.10 (95% CI: 0.60–1.00) and an accuracy of 0.88 ± 0.04 (95% CI: 0.83–0.94).
Class-Specific Metrics: The model demonstrated high specificity and perfect recall for the negative class (non-toxicity). However, sensitivity for the minority positive class (toxicity) was lower (0.33), resulting in an F1-score of 0.50. The authors note this is consistent with the challenge of detecting rare events in small, imbalanced cohorts, though the high precision (1.00) suggests that positive predictions are reliable.
Interpretability: SHAP analysis identified a compact set of plausible predictors:
- Anatomical Geometry: The bladder's maximum 2D diameter (row direction) was the most influential feature; larger dimensions correlated with higher risk. Higher prostate flatness was associated with lower risk, while a longer rectum major axis increased risk.
- Dose Patterns: Higher bladder GLRLM normalized run-length non-uniformity (indicating fragmented, heterogeneous dose patterns) increased risk. Conversely, lower intraprostatic first-order Interquartile Range (IQR) was associated with higher risk, suggesting that reduced dose variability within the target might reflect less optimal SBRT configurations.

Significance and Claims
The authors position this study as one of the first to specifically investigate a dosiomics-driven predictive model for late GU toxicity in the context of CyberKnife SBRT for prostate cancer. The primary significance lies in the combination of:

Predictive Capability: Demonstrating that dosiomic features (capturing dose heterogeneity and spatial organization) can discriminate late toxicity outcomes better than traditional DVH-only approaches might suggest.
Clinical Interpretability: Moving beyond "black box" predictions by using SHAP to provide clinically interpretable risk drivers. The study identifies that both organ geometry (e.g., bladder distension) and dose-pattern heterogeneity (e.g., fragmented dose in the bladder) are critical factors.
Workflow Integration: Proposing a framework that supports transparent risk stratification and potential planning optimization.

Limitations and Future Directions
The paper maintains a modest tone regarding its claims, explicitly acknowledging limitations:

Sample Size: The small cohort (87 patients) and low prevalence of the endpoint limit statistical power and increase the risk of residual overfitting.
Generalizability: As a single-institution study, variations in imaging, delineation, and planning across centers may affect model transportability.
Sensitivity: The lower sensitivity for the toxicity class indicates a need for larger datasets to improve detection of rare adverse events.

Consequently, the authors conclude that while the results are encouraging, multicenter validation on larger datasets is required to confirm generalizability before clinical adoption. They do not claim the model is ready for immediate clinical deployment but rather as a proof-of-concept for a dosiomics-driven, explainable approach to treatment personalization.