Deep learning (DL) models have demonstrated state-of-the-art performance in the classification of diagnostic imaging in oncology. However, DL models for medical images can be compromised by adversarial images, where pixel values of input images are manipulated to deceive the DL model. To address this limitation, our study investigates the detectability of adversarial images in oncology using multiple detection schemes. Experiments were conducted on thoracic computed tomography (CT) scans, mammography, and brain magnetic resonance imaging (MRI). For each dataset we trained a convolutional neural network to classify the presence or absence of malignancy. We trained five DL and machine learning (ML)-based detection models and tested their performance in detecting adversarial images. Adversarial images generated using projected gradient descent (PGD) with a perturbation size of 0.004 were detected by the ResNet detection model with an accuracy of 100% for CT, 100% for mammogram, and 90.0% for MRI. Overall, adversarial images were detected with high accuracy in settings where adversarial perturbation was above set thresholds. Adversarial detection should be considered alongside adversarial training as a defense technique to protect DL models for cancer imaging classification from the threat of adversarial images.

Purpose or Objective: To evaluate the feasibility and clinical utility of integrating sequential PSMA-PET imaging into an offline–online adaptive workflow for response-based dominant intraprostatic lesion (DIL)-boosting high-risk prostate cancer treated with stereotactic ablative radiotherapy (SABR). Materials and Methods: As part of a prospective trial, patients were treated on MR- or CBCT-guided adaptive radiotherapy (ART) systems with prostate/pelvic node 5-fraction SABR (36.25 Gy/25 Gy) with DIL boost (50 Gy). Whereas traditional DIL boost volumes delineate full pre-therapy imaging-defined disease (GTVinitial), this study serially refined DIL boost volumes based on treatment response defined by PSMA-PET scans after neoadjuvant androgen deprivation therapy (nADT, GTVmb1) and fraction 3 SABR (GTVmb2). DIL delineation employed PET-PSMA fusion to CT/MR simulation and was guided by a rule-based %SUVmax threshold approach. Comparisons of GTV volumes and OAR dosimetry were performed between plans using GTVinitial versus GTVmb1/GTVmb2 for DIL boost, for each of the initial cohorts of five patients from the initially treated cohorts. Results: Five patients treated on MR-Linac (n = 3) or CBCT-based ART (n = 2) were analyzed. Three patients exhibited complete imaging response after nADT, omitting GTVmb boosts. Offline GTVmb refinements based on PSMA-PET were seamlessly integrated into ART workflows without introducing additional treatment time. DIL GTV volumes significantly decreased (p = 0.03) from an initial mean of 11.4 cc (GTVinitial) to 4.1 cc (GTVmb1) and 3.0 cc (GTVmb2). Dosimetric analysis showed meaningful reductions in OAR doses: rectal wall D0.035 cc decreased by up to 12 Gy, while bladder wall D0.035 cc and V18.3 Gy reduced from 52.3 Gy and 52.3 cc (Plan_initial) to 42.9 Gy and 24.9 cc (Plan_mb2), respectively. Urethra doses remained stable, with minor reductions. Sigmoid and femoral head doses remained within acceptable limits. Online adaptation efficiently addressed daily anatomical variations, enabling simulation-free plan re-optimization. Conclusion: PSMA-PET-guided adaptive microboosting for HRPCa SABR is feasible and effective. Standard MR-Linac and CBCT systems offer practical alternatives to BgRT platforms, enabling biology-driven dose personalization and potentially reducing toxicity.

PSMA PET/CT imaging has been increasingly utilized in the management of patients with metastatic prostate cancer (mPCa). Imaging biomarkers derived from PSMA PET may provide improved prognostication and prediction of treatment response for mPCa patients. This study investigates a novel deep learning-derived imaging biomarker framework for outcome prediction using multi-modal PSMA PET/CT and clinical features. A single institution cohort of 99 mPCa patients with 396 lesions was evaluated. Imaging features were extracted from cropped lesion areas and combined with clinical variables including body mass index, ECOG performance status, prostate specific antigen (PSA) level, Gleason score, and treatments received. The PSA progression-free survival (PFS) model was trained using a ResNet architecture with a Cox proportional hazards loss function using five-fold cross-validation. Performance was assessed using concordance index (C-index) and Kaplan-Meier survival analysis. Among evaluated model architectures, the ResNet-18 backbone offered the best performance. The multi-modal deep learning framework achieved a 5-fold cross-validation C-index ranging from 0.75 to 0.94, outperforming models incorporating imaging only (0.70–0.89) and clinical features only (0.53–0.65). Kaplan-Meir survival analysis performed on the deep learning-derived predictions demonstrated clear risk stratification, with a median PSA progression free survival (PFS) of 19.7 months in the high-risk group and 26 months in the low-risk group ( P < 0.001). Deep learning-derived imaging biomarker based on PSMA PET/CT can effectively predict PSA PFS for mPCa patients. Further clinical validation in prospective cohorts is warranted.

Importance: Artificial intelligence (AI) foundation models such as Segment Anything Model 2 (SAM 2) offer potential for semi-automated image segmentation with minimal fine-tuning, but their performance in specialized clinical tasks like radiation therapy planning are not well characterized. Objective: To evaluate the performance of SAM 2 in segmenting pre-operative intact prostate and post-operative prostate fossa targets for prostate radiotherapy planning. Design, Setting, Participants: Retrospective cohort study deploying and testing a foundation model for AI segmentation for prostate radiotherapy planning. CT simulation images and radiation plans were obtained from a single academic institution for patients undergoing prostate cancer treatment. Data analysis was performed from September 2024 to February 2025. Exposures: AI segmentation with varying levels of human intervention, ranging from intervals of every 2nd to every 10th ground truth slice provided as input. Main Outcome and Measures: Segmentation accuracy measured by Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) for intact and post-operative prostate target delineation. Results: While SAM 2 outperformed interpolation in DSC and HD for both intact and post-operative prostate cancer patient cases, the AI segmentation accuracy was significantly better in the intact pre-operative patient cases where anatomic boundaries were better defined than post-operative patient cases. This is especially evident when sparse ground truth was provided simulating lower levels of human intervention. Conclusions and Relevance: AI foundation models show promising application for specialized medical tasks such as prostate cancer radiotherapy segmentation with limited need for fine-tuning or retraining, although their clinical application will require further understanding of task-specific performance.

Our study evaluated whether a deep learning auto segmentation model combined with machine learning triage can streamline radiotherapy clinical trial quality assurance (QA). We analyzed 107 stereotactic ablative radiotherapy (SABR) cases from a multi-institutional phase II clinical trial of neurovascular sparing prostate SABR, focusing on physician contours of the internal pudendal artery (IPA) as a novel organ-at-risk with substantial interobserver variability. Contours were scored by the trial principal investigator as Per-Protocol or Minor Deviation/Unacceptable. We applied a deep learning model for IPA auto-segmentation. Agreement between human and AI contours was then quantified using 14 overlap, distance, and surface metrics, and a supervised classifier was trained on these metrics to flag clinical trial protocol deviations. While AI segmentation achieved only modest geometric accuracy with mean Dice similarity coefficient of 0.446 and 95th percentile Hausdorff distance of 14.23, when incorporating all 14 metrics, a machine learning classifier yielded AUROC of 0.836, flagging all Minor Deviation/Unacceptable cases with 100% sensitivity on the 27 case hold-out set with 6 false positives and no false negatives. AI segmentation combined with metrics-based machine learning can triage protocol deviations within a multi-institution radiotherapy clinical trial, supporting prospective evaluation of AI-assisted trial QA.

Our study evaluated whether a deep learning auto-segmentation model combined with machine learning triage could support radiotherapy clinical trial quality assurance (QA). We analyzed 107 stereotactic ablative radiotherapy (SABR) cases from a multi-institutional phase II clinical trial of neurovascular sparing prostate SABR, focusing on physician contours of the internal pudendal artery (IPA) as a novel organ-at-risk with substantial interobserver variability. Contours were scored by the trial principal investigator as Per-Protocol (n = 89) or Minor Deviation/Unacceptable (n = 18). We applied a deep learning model for IPA auto-segmentation. Agreement between human and AI contours was then quantified using 14 overlap, distance, and surface metrics, and a supervised classifier was trained on these metrics to flag clinical trial protocol deviations. While AI segmentation achieved only modest geometric agreement with mean Dice similarity coefficient of 0.446 and 95th percentile Hausdorff distance of 14.23, when incorporating all 14 metrics, a machine learning classifier yielded an AUROC of 0.836 (95% CI, 0.552–0.949), flagging all Minor Deviation/Unacceptable cases with 100% sensitivity (95% CI, 56.6–100%) on the 27 case hold-out set with 6 false positives and no false negatives. AI segmentation combined with metrics-based machine learning can triage protocol deviations within a multi-institutional radiotherapy clinical trial, supporting prospective evaluation of AI-assisted trial QA. However, the small number of hold-out protocol deviation cases in this analysis limits generalizability, and the clinical acceptability of the AI contours and potential reductions in QA workload require further investigation.