Purpose: Interfractional anatomical changes reduce the accuracy of radiotherapy treatment. However, developing new plans to account for daily changes is typically impractical. We propose a U-Net model to predict the daily dose distribution, enabling dose deviation evaluation to inform decisions regarding the necessity of re-planning.

 

Methods: A U-Net was trained using data from 13 patients receiving intensity modulated radiotherapy (IMRT) (63 Gy/30 fx) for advanced lung cancer. The training dataset included planning CT and daily cone-beam CT (CBCT) images, along with their associated dose distributions. Rigid registration was performed to align CBCT images with the corresponding planning CT images, from which dose was re-calculated in the treatment planning system. These CBCT-based dose distributions were normalized and assumed to represent the daily delivered, relative dose. Leave-one-out cross-validation method was employed to produce 13 separate networks, where one patient dataset was excluded per model and used for validation. Model accuracy was quantified using gamma analysis with 3%/3mm and 2%/2mm criteria, along with a 20% maximum dose threshold.

 

Results: The models yielded mean gamma pass rates of 98.4 ± 1.3% and 94.7 ± 2.4% for 3%/3mm and 2%/2mm tolerances, respectively. Gamma failures within the treated volume were randomly distributed, showing no apparent correlation with anatomical structures, beam edges, or dose gradients.

 

Conclusions: Model predictions exhibited strong agreement with daily delivered dose distributions, likely due to the relatively low geometric variability in the training dataset. The next step is to test robustness of the models in predicting absolute doses with increased interfractional variability.

We deployed a locally hosted, open-source large language model (Llama3.3-70B) at our tertiary cancer center to automatically extract structured prognostic information from 1,360 unstructured head and neck oncology notes (pathology and consultation). Installed on two NVIDIA GPUs (32 Gigabytes each) within our hospital, the model processes patient records internally, preserving privacy. We targeted 30 key expert-guided clinical and pathological parameters, including sex, anatomical lesion site, pathological and clinical TNM staging, p16/human papillomavirus (HPV) status, immunohistochemical profiles (e.g., EBER), Charlson Comorbidity Scores, performance statuses (Karnofsky Performance Scale and the Eastern Cooperative Oncology Group Performance Status), medication history, symptoms at presentation, and detailed lifestyle factors such as smoking (packs per year) and alcohol consumption. We combined few-shot learning, human-in-the-loop prompt refinements, and chain-of-thought reasoning for complex fields requiring inference beyond explicit text. A random sample of 30 reports (900 total extraction points) was retrospectively evaluated by a senior radiation oncologist, achieving near 100% accuracy, precision, recall, and F1 scores. On average, extraction time fell from 6–10 minutes per note (manual reviewer) to 30–60 seconds per note (1–2 seconds per extraction), representing an approximate tenfold efficiency gain. Moreover, this pipeline uncovered underutilized prognostic indicators, such as resection margin, perineural and lymphovascular invasion findings or comorbidity burden, emphasizing its value for clinical decision support. In conclusion, our pipeline demonstrates the feasibility of a locally deployed large language model in head and neck oncology, offering scalable, privacy compliant data extraction to catalyze large-scale research, tumor board clinical support, and personalized treatment planning.

Purpose: To develop a principal component analysis (PCA)-based future frame prediction method for non-invasive intra-fractional tumour-tracked radiotherapy (nifteRT) using a hybrid Linac-MR system.

 

Methods: Sagittal dynamic images from 10 lung and 10 liver patients were acquired using a 3T MRI with a 2D bSSFP protocol at 4 frames/s. A window of past 60 frames was selected to characterize recent variations using PCA. PCA was applied in k-space, and the 16 most significant principal components (PCs) were used for prediction. The temporal trajectory formed by the PC scores was extrapolated independently to 1 and 2 frames beyond the window, using auto-regression. Using the 16 extrapolated scores as weights, the PCs were combined to generate predictive images. Image-based metrics including structural similarity (SSIM) were used to assess accuracy. The same auto-contouring technique was applied to both predicted and acquired images, and the consistency of tumour position and shape was measured using Dice coefficient (DC).

 

Results: We report an average SSIM of 0.93 for 1-frame, and 0.92 for 2-frame predictions for both liver and lung. The contour analysis yielded agreement with corresponding acquired images for 1- and 2-frame predictions, with average DC of 0.94 and 0.90 for liver, and 0.91 and 0.88 for lung. Average prediction time was 49 ms/frame.

 

Conclusions: The presented methodology offers a computationally efficient and accurate means of predicting images in real-time for nifteRT on a Linac-MR. The predicted contours have a sufficient level of consistency to consider for use in directing MLC motions in advance of anatomic motion.

Purpose: To investigate the effects of slice thicknesses on radiomics feature extraction for kidney lesion differentiation, highlighting the role of harmonization in reducing bias.

 

Methods: This study used the KITS23 dataset comprising CT images of 599 patients, including 334 (198 tumors, 136 cysts) with 5 mm and 62 (45 tumors, 17 cysts) with 0.5 mm slice thickness. Classifiers were trained on 60% of balanced datasets of 0.5 mm and 5 mm slice thicknesses, while a third dataset combining mixed slice thicknesses underwent harmonization using NestedCombat. Radiomic features were extracted using PyRadiomics from CT images from annotated lesions. Feature reduction was conducted using LASSO with 10-fold cross-validation. Models were developed with the Relief feature selection method across ten ML algorithms, incrementally adding features. Hyperparameters were refined through 10-fold cross-validation in GridSearchCV, and model performance was assessed by the AUC on their respective test datasets.

 

Results: Models trained on the 0.5 mm slices consistently outperformed those on the 5.0 mm slices. Bagging, an ensemble method, recorded the highest performance gap with a 14% difference in AUC between thin and thick slices. Harmonization preserved radiomic feature discrimination, resulting in a consistent median AUC of 0.91 for the mixed dataset, compared to 0.92 for 0.5 mm and 0.86 for 5.0 mm.

 

Conclusion: This study emphasizes the impact of slice thickness on radiomic analysis, showing that thinner slices enhance classifier accuracy, commensurate with the expected larger information content of these images. Harmonization minimizes performance discrepancies, illustrating its crucial role in standardizing radiomic assessments.

Purpose: Automatic treatment planning tools, like RapidPlan (RP) within the Eclipse treatment planning system, leverage machine learning to predict optimal dose objectives. However, the effectiveness of RP is heavily reliant on the availability of high-quality training data. This study presents a novel method for generating high-quality training data using synthetic dose distributions.

 

Materials and Methods: A MATLAB-based tool, Synthetic-Dose-Generator, was developed to simulate dose distributions using a double exponential model. These generated dose distributions are entirely determined by the geometry of the targets and organs-at-risk (OARs). The tool was employed to create synthetic dose distributions for 25 head-and-neck (HN) cases (70Gy/56Gy/35 fractions) and 25 prostate cases (40Gy/36.25Gy/5 fractions). Using the generated synthetic dose distributions, two synthetic RP models (synRP) were created for each site and compared to clinical RP models (clRP) from 171 HN cases and 97 prostate cases. Plan quality was assessed based on target coverage, OAR doses, and monitor units (MUs), while deliverability was evaluated through portal dosimetry (3%/2mm).

 

Results: Target coverage between synRP and clRP was identical. synRP significantly reduced OAR doses: for prostate, bladder and rectum doses dropped by 7% (p<0.005) and 17% (p<0.0005), respectively. In HN cases, synRP improved the sparing of parallel OARs by 6% (p<0.00005) and serial OARs by 4% (p<0.005). synRP plans required slightly more MUs (618±60 vs 595±68 for HN, 2922±166 vs 2872±228 for prostate). Both plans passed gamma evaluation (>98%).

 

Conclusion: The synRP model enhances plan quality, effectively generating high-quality RP models with fewer clinical cases and reduced reliance on historical data.

While deep learning has revolutionized digital pathology, most models rarely integrate clinical variables into foundation pathology models, especially for TCGA-HNSC (head and neck squamous cell carcinoma), a cohort known for poor outcome prediction and often excluded from benchmarks. To address this, we systematically evaluated survival modeling in TCGA-HNSC by fusing whole-slide image (WSI) embeddings from the Hierarchical Image Pyramid Transformer with clinical data from the Feature Tokenizer Transformer, applying a Cox Proportional Hazards model for individualized risk estimation. We explored three fusion paradigms: marginal fusion, where embeddings were concatenated before training; joint fusion, leveraging cross-attention and variational autoencoders (VAEs); and late fusion, which applied meta-learning to combine modality-specific models. With five-fold cross-validation, and C-index as evaluation metric, it was shown that WSIs alone achieved a C-index of 0.52, while clinical data alone reached 0.63. Joint fusion failed to surpass clinical-only models (C-index ~0.63), while marginal fusion provided a small gain (0.64), showing that direct cross-modal interaction had limited benefit. However, late fusion with meta-learning achieved 0.66, surpassing all baselines. Additionally, our framework enabled risk stratification, effectively separating patients into high- and low-risk groups (p < 0.05), enhancing clinical interpretability. Our study is the first to systematically evaluate fusion strategies for TCGA-HNSC survival prediction, pioneering multimodal integration of histopathology and clinical data. These findings underscore the importance of tailored fusion rather than assuming any multimodal combination improves predictive power, establishing a benchmark for future survival modeling approaches.