Purpose. Different ML models were compared to predict toxicity in RT on a large cohort (n = 1314). Methods. The endpoint was RTOG G2/G3 acute toxicity, resulting in 204/1314 patients with the event. The dataset, including 25 clinical, anatomical, and dosimetric features, was split into 984 for training and 330 for internal tests. The dataset was standardized; features with a high p-value at univariate LR and with Spearman $\rho >0.8$ were excluded; synthesized data of the minority were generated to compensate for class imbalance. Twelve ML methods were considered. Model optimization and sequential backward selection were run to choose the best models with a parsimonious feature number. Finally, feature importance was derived for every model. Results. The model’s performance was compared on a training–test dataset over different metrics: the best performance model was LightGBM. Logistic regression with three variables (LR3) selected via bootstrapping showed performances similar to the best-performing models. The AUC of test data is slightly above 0.65 for the best models (highest value: 0.662 with LightGBM). Conclusions. No model performed the best for all metrics: more complex ML models had better performances; however, models with just three features showed performances comparable to the best models using many (n = 13–19) features.

Background/Objectives: Pancreatic cancer is a very aggressive disease with a poor prognosis, even when diagnosed at an early stage. This study aimed to validate and refine a radiomic-based [¹⁸F]FDG-PET model to predict distant relapse-free survival (DRFS) in patients with unresectable locally advanced pancreatic cancer (LAPC). Methods: A Cox regression model incorporating two radiomic features (RFs) and cancer stage (III vs. IV) was temporally validated using a larger cohort (215 patients treated between 2005–2022). Patients received concurrent chemoradiotherapy with capecitabine and hypo-fractionated Intensity Modulated Radiotherapy (IMRT). Data were split into training (145 patients, 2005–2017) and validation (70 patients, 2017–2022) groups. Seventy-eight RFs were extracted, harmonized, and analyzed using machine learning to develop refined models. Results: The model incorporating Statistical-Percentile10, Morphological-ComShift, and stage demonstrated moderate predictive accuracy (training: C-index = 0.632; validation: C-index = 0.590). When simplified to include only Statistical-Percentile10, performance improved slightly in the validation group (C-index = 0.601). Adding GLSZM3D-grayLevelVariance to Statistical-Percentile10, while excluding Morphological-ComShift, further enhanced accuracy (training: C-index = 0.654; validation: C-index = 0.623). Despite these refinements, all versions showed similar moderate ability to stratify patients into risk classes. Conclusions: [¹⁸F]FDG-PET radiomic features are robust predictors of DRFS after chemoradiotherapy in LAPC. Despite moderate performance, these models hold promise for patient risk stratification. Further validation with external cohorts is ongoing.

In this proceedings we demonstrate some advantages of a top-bottom approach in the development of hardware-accelerated code. We start with an autogenerated hardware-agnostic Monte Carlo generator, which is parallelized in the event axis. This allow us to take advantage of the parallelizable nature of Monte Carlo integrals even if we don't have control of the hardware in which the computation will run (i.e., an external cluster). The generic nature of such an implementation can introduce spurious bottlenecks or overheads. Fortunately, said bottlenecks are usually restricted to a subset of operations and not to the whole vectorized program. By identifying the more critical parts of the calculation one can get very efficient code and at the same time minimize the amount of hardware-specific code that needs to be written. We show benchmarks demonstrating how simply reducing the memory footprint of the calculation can increase the performance of a $2 \to 4$ process.

Purpose: This paper aimed to identify dosimetric, clinical, and CT-based densitometric predictors of radiation-induced pulmonary events in breast cancer patients treated with moderately hypofractionated radiotherapy. Materials and Methods: A single-institution cohort of 1172 consecutive patients treated with 3D conformal whole-breast radiotherapy (40 Gy/15 fractions) before 2017 was analyzed. Ipsilateral lung DVHs and CT densitometry metrics were extracted. Clinical variables and cardiac calcification (CAC) scores (Agatston_score, CAC_volume, Max_HU_Heart) were included. Univariable and multivariable logistic regressions were performed; collinearity was assessed via Spearman correlation and VIF. Optimal thresholds were derived using the Youden index. Internal validation used bootstrap resampling. Results: After a median follow-up of 6.5 years, 18 patients developed moderate/severe pulmonary events. The univariable analysis showed associations with lung densitometric features (median/mean HU, 10th percentile, the lung volume with HU < −850 (V850)), V37 Gy, lung volume, and CAC scores. Lower lung HU values and larger lung volumes were linked to higher risk. The best models combined V850 (or lung volume) with a CAC metric. The model including V850 > 175 cc and continuous Max_HU_Heart achieved an optimism-corrected AUC of 0.68, with good fit and calibration (Hosmer–Lemeshow p = 0.33, R² = 0.847). Conclusions: The baseline cardiopulmonary status, captured by lung and heart densitometry, predicts pulmonary toxicity better than dosimetry. V850 > 175 cc was associated with a 4-fold higher risk, consistent with air trapping, known as a marker of emphysema.

In this proceedings we demonstrate some advantages of a top-bottom approach in the development of hardware-accelerated code. We start with an autogenerated hardware-agnostic Monte Carlo generator, which is parallelized in the event axis. This allow us to take advantage of the parallelizable nature of Monte Carlo integrals even if we don't have control of the hardware in which the computation will run (i.e., an external cluster). The generic nature of such an implementation can introduce spurious bottlenecks or overheads. Fortunately, said bottlenecks are usually restricted to a subset of operations and not to the whole vectorized program. By identifying the more critical parts of the calculation one can get very efficient code and at the same time minimize the amount of hardware-specific code that needs to be written. We show benchmarks demonstrating how simply reducing the memory footprint of the calculation can increase the performance of a $2 \to 4$ process.