PURPOSEIt is essential to detect and segment liver tumors to guide treatment and track disease progression. To reduce the need for large annotated data sets, we present an end-to-end pipeline that uses self-supervised pretraining to improve segmentation and then classifies tumor types with a separate pretrained classifier applied to the segmented tumor regions.METHODSFirst, we pretrained the encoder of a transformer-based network using a self-supervised approach on unlabeled abdominal computed tomography images. Subsequently, we fine-tuned the segmentation network to segment the liver and tumors, and the tumor regions were classified using a pretrained convolutional neural network (Inception-v3 architecture) as intrahepatic cholangiocarcinoma (ICC), hepatocellular carcinoma (HCC), or colorectal liver metastases (CRLMs). We evaluated 459 images (155 HCC, 107 ICC, 197 CRLM). For external testing, we used an independent public data set (n = 40).RESULTSAveraged across HCC, ICC, and CRLM, in comparison with a supervised baseline (no pretraining), self-supervised pretraining improved the liver Dice similarity coefficient (DSC) by 6.4 percentage points and reduced the 95th-percentile Hausdorff distance (HD95) by 32.97 mm. For tumors, the DSC increased by 6.0 percentage points and the HD95 decreased by 3.2 mm. Tumor type classification achieved AUC 0.98 (95% CI, 0.96 to 1.00) and accuracy 96% (95% CI, 92% to 99%). Segmentation performance on the external data was close to the internal cohort with tumor DSC 0.73, intersection over union (IoU) 0.60, and HD95 30.98 mm and liver DSC 0.91, IoU 0.83, and HD95 29.67 mm.CONCLUSIONThe proposed self-supervised, end-to-end pipeline improves liver tumor segmentation and provides accurate tumor type classification, supporting reliable radiologic assessment, treatment planning, and improved prognostication for patients with liver cancer.

BACKGROUND:Radiomic imaging biomarkers are increasingly studied in oncology as a means to support disease prognosis and personalized treatment planning. While deep learning (DL) offers scalable alternatives to handcrafted radiomic features, DL-derived biomarkers are sensitive to variations in image acquisition protocols and scanner hardware-even within a single imaging modality. To ensure reliable and reproducible biomarker estimation, it is essential to (1) provide clinicians with quantitative uncertainty estimates associated with biomarker predictions, and (2) address acquisition-induced variability through harmonization strategies that render consistent performance across diverse imaging conditions. PURPOSE:To evaluate uncertainty estimation as a reliability metric for biomarker prediction and to assess the role of image harmonization in improving cross-scanner inference, we develop deep learning radiomics (DLR) models for joint estimation of biomarkers and associated uncertainties and apply linear harmonization filters to standardize imaging conditions. METHODS:We constructed hybrid digital phantoms by embedding 20,000 virtual colorectal liver metastases into 20 clinical CT liver images to generate metastases-laden simulated scans. The proposed DLR models jointly estimated a selected set of biomarkers and their associated aleatoric uncertainties from simulated images. Models were trained, validated, and tested using a 70:20:10 data split. Variability in CT image acquisition was modelled using two scanner types, two reconstruction kernels, and three x-ray tube current settings. DLR performance was evaluated under direct inference (matched training and test scanners), cross-scanner inference, and harmonized inference using linear harmonization filters. Quantitative evaluation was based on correlation coefficients and root mean squared errors (RMSEs). RESULTS:The estimated uncertainties were consistently higher for biomarker predictions with larger deviations from ground truth. Excluding high-uncertainty predictions improved concordance between predictions and ground truth biomarkers. The proportion of uncertain predictions increased under cross-scanner inference relative to direct inference, indicating reduced biomarker reliability under heterogeneous imaging conditions. Furthermore, application of harmonization filters reduced RMSEs by an average of 43% across biomarkers and experiments during cross-scanner inference, demonstrating improved cross-scanner consistency. CONCLUSIONS:In this controlled in-silico study, uncertainty estimation provided a practical reliability metric for DL-based radiomic biomarker prediction, while image harmonization improved reproducibility across heterogeneous acquisition conditions. These findings demonstrate methodological feasibility within a simulation-based framework but have not yet been validated on clinical CT data, motivating future clinical validation and translation.

Artificial intelligence (AI) in healthcare has led to many promising developments; however, increasingly, AI research is funded by the private sector leading to potential trade-offs between benefits to patients and benefits to industry. Health AI practitioners should prioritize successful adaptation into clinical practice in order to provide meaningful benefits to patients, but translation usually requires collaboration with industry. We discuss three features of AI studies that hamper the integration of AI into clinical practice from the perspective of researchers and clinicians. These include lack of clinically relevant metrics, lack of clinical trials and longitudinal studies to validate results, and lack of patient and physician involvement in the development process. For partnerships between industry and health research to be sustainable, a balance must be established between patient and industry benefit. We propose three approaches for addressing this gap: improved transparency and explainability of AI models, fostering relationships with industry partners that have a reputation for centering patient benefit in their practices, and prioritization of overall healthcare benefits. With these priorities, we can sooner realize meaningful AI technologies used by clinicians where mutually beneficial impacts for patients, healthcare providers, and industry can be realized.

Reliable prognostic models of death or liver recurrence following resection of colorectal liver metastases are critical to stratify patients for treatment. The study aimed to develop models incorporating clinical and imaging data into multimodal preoperative prediction models of hepatic disease-free survival and overall survival. We conducted a retrospective cohort study with 1,301 consecutive patients from Memorial Sloan Kettering Cancer Center and The University of Texas MD Anderson Cancer Center. Clinical and computed tomography (CT) data were included in radiomic and deep learning models and compared. Our findings suggest that a deep learning model that utilizes the tumor region from CT imaging is the highest performing individual model with C-index of 0.61 for both hepatic disease-free (HDFS) and overall survival. Combining radiomic and clinical models into a multimodal 'clinicoradiomic' model (C-index = 0.63 [0.58-0.69]) outperformed unimodal models (C-index = 0.61 [0.55-0.66]) and the current clinical risk score (C-index = 0.55 [0.49-0.60]) for HDFS. Our study successfully developed preoperative imaging models that integrate clinical data to predict patients at risk of hepatic recurrence, outperforming all currently available models.

In medical image analysis, feature engineering plays an important role in the design and performance of machine learning models. Persistent homology (PH), from the field of topological data analysis (TDA), demonstrates robustness and stability to data perturbations and addresses the limitation from traditional feature extraction approaches where a small change in input results in a large change in feature representation. Using PH, we store persistent topological and geometrical features in the form of the persistence barcode whereby large bars represent global topological features and small bars encapsulate geometrical information of the data. When multiple barcodes are computed from 2D or 3D medical images, two approaches can be used to construct the final topological feature vector in each dimension: aggregating persistence barcodes followed by featurization or concatenating topological feature vectors derived from each barcode. In this study, we conduct a comprehensive analysis across diverse medical imaging datasets to compare the effects of the two aforementioned approaches on the performance of classification models. The results of this analysis indicate that feature concatenation preserves detailed topological information from individual barcodes, yields better classification performance and is therefore a preferred approach when conducting similar experiments.

Background and Objective: This study introduces the liver cancer segmentator (LCS), a deep learning model designed for automatic and robust segmentation of liver parenchyma and tumors in abdominal contrast-enhanced computed tomography images from patients with colorectal liver metastases. The primary aim was to enhance confidence scoring for more reliable clinical segmentation assessment. Methods: In this retrospective study, 446 abdominal contrast-enhanced computed tomography examinations were collected; 355 (80%) were used for training and 91 for testing. Data originated from routine clinical cases at two institutions, representing diverse disease stages and treatment settings. A state-of-the-art neural network segmentation framework was trained on these cases, with performance evaluated using the Dice score and the normalized surface distance. An iterative training process, supported by an integrated annotation workflow, was employed to refine the training set. The final model was applied to the 91 test examinations to assess the impact of tumor volume and slice thickness on confidence scoring. Reliability was quantified through pairwise Dice score for failure detection and the area under the risk coverage curve. Results: The LCS achieved a Dice score of 0.9707 (95% CI: 0.9663-0.9751) for liver parenchyma and 0.7695 (95% CI: 0.7166-0.8224) for tumors. Normalized surface distance values at a 3-millimeter tolerance were 0.9605 (95% CI: 0.9539-0.9671) for parenchyma and 0.8412 (95% CI: 0.7928-0.8896) for tumors. Confidence scoring analysis demonstrated strong correlations between tumor volume, slice thickness, and segmentation reliability, reducing the area under the risk coverage curve from 16.7 to 10.3. Conclusions: The LCS achieved high segmentation accuracy in patients with colorectal liver metastases. Incorporating tumor volume and slice thickness into the confidence scoring process improved failure detection, enhanced reliability, and provided valuable insights for refining clinical deployment of automated segmentation algorithms.

OBJECTIVE:Carotid plaque detected by ultrasound is associated with major adverse cardiovascular events (MACE) and can be characterized using manual or automated radiomic analysis. The generated plaque characteristics may have complex interdependencies for which machine learning (ML) may provide predictive modeling. The objective of the study was to develop a ML model to predict MACEs from clinical variables, manual focused vascular ultrasound (FOVUS) measurements, and semi-automated radiomic ultrasound features. METHODS:Carotid ultrasound scans were performed on 493 patients and MACE outcomes were collected over 5 y by medical chart review. ML-based models were compared that incorporated clinical characteristics, manual FOVUS measurements, and quantitative radiomic features. Feature selection was performed via ReliefF, with a sample:feature ratio of 10:1 yielding 11 top features for training the ML model. Four ML classifiers were developed using 10 K-fold cross-validation. RESULTS:Over 5 years, 144 patients (29%) experienced a MACE outcome (death, unstable angina, myocardial infarction, stroke, transient ischemic attack, carotid endarterectomy, coronary artery bypass graft, or percutaneous coronary intervention). The best model (x-gradient boost) performed significantly better than chance alone (p = 2 × 10-7), with the highest average prediction accuracies (clinical data only: 0.849 ± 0.039, FOVUS only: 0.875 ± 0.029, radiomics only: 0.705 ± 0.036, clinical and FOVUS data: 0.885 ± 0.023, clinical and radiomics data: 0.804 ± 0.053, FOVUS and radiomic data: 0.935 ± 0.043, all data: 0.958 ± 0.023). CONCLUSION:The combination of clinical, FOVUS and radiomics data provides strong predictive power for distinguishing MACE from non-MACE. Moreover, both automated and manual plaque quantification can generate predictive features.

State-of-the-art interactive segmentation with foundation models remains constrained by compute and the need for annotations. We evaluated parameter-efficient fine-tuning (PEFT) of the Segment Anything Model (SAM) by inserting baseline adapters, Low-Rank Adaptation (LoRA), 4-bit Quantized Low-Rank Adaptation (QLoRA), and a convolutional adapter (Conv-Adapter), as well as our proposed Directional Spectral Top-K adapter (DiSCo), training only the adapters while keeping the SAM backbone frozen. DiSCo performs singular value decomposition of row-normalized weights to obtain spectral bases, then learns rank-gated spectral coefficients, per-output magnitude offsets, and a spectral gain; it supports optional Top-K rank selection at inference while keeping only 0.14M parameters trainable. We benchmarked five prompting regimes, none, single-point, multi-point, and bounding boxes at an intersection over union (IoU) of 0.50 and 0.75, applied to abdominal CT scans of colorectal liver metastases, and reported segmentation metrics (Dice and the 95th-percentile Hausdorff distance (HD95)) and compute metrics (trainable parameters, latency, memory, throughput). Conv-Adapter and LoRA achieved the highest accuracy (overall Dice 0.793 and 0.792; single-point 0.795 and 0.792; HD95 32 mm). QLoRA was close (overall 0.766; single-point 0.768; HD95 36.41mm) while offering the most favorable compute profile (0.91M trainable parameters, 120 ms latency, 4.9GB peak memory). DiSCo maximized parameter efficiency, achieving the highest Dice per million trainable parameters (4.66), showing an accuracy-efficiency trade-off (overall Dice 0.653; single-point 0.698; HD95 49.53 mm). These results indicate that PEFT on foundation models enables more accurate liver tumor segmentation with reduced adaptation costs, supporting rapid scanner-specific tuning, point- or box-prompt workflows, and broader deployability for preoperative volumetrics when compute and labeled data are limited. The code supporting this study is available at: https://github.com/RamtinMojtahedi/PEFT-SAM-Liver-CT.

Accurate segmentation of colorectal liver metastases (CRLM) in contrast-enhanced computed tomography (CT) is important for response assessment, surgical planning, and follow-up. We propose two parameter-efficient spectral adapters for the Segment Anything Model (SAM): the Directional Spectral Adapter (DiSECT) and Spectral Instance-Guided Adapter (SiGA). DiSECT uses singular value decomposition of frozen weights to constrain residual updates to leading spectral directions, while SiGA adds global and input-conditioned gating through a multilayer perceptron. We evaluate these methods on 446 contrast-enhanced CT volumes (355 training, 91 testing) and compare them with LoRA, QLoRA, convolutional adapters (CAD), and a 3D nnU-Net baseline. Experiments consider single-point, three-point, bounding-box, and no-prompt regimes. SiGA achieves the best single-point performance with a Dice score of 0.77, IoU of 0.69, and HD95 of 35.39 mm. Under no-prompt inference, SiGA reaches 0.76 Dice, 0.68 IoU, and 46.76 mm HD95, comparable to the nnU-Net baseline (0.758 Dice). DiSECT uses only 0.14 million trainable parameters. These results show that spectral adapters can efficiently adapt SAM for CRLM segmentation while retaining strong accuracy with limited trainable parameters.

Purpose Colorectal cancer is the third most common cancer globally, with a high mortality rate due to metastatic progression, particularly in the liver. Surgical resection remains the main curative treatment, but only a small subset of patients is eligible for surgery at diagnosis. For patients with initially unresectable colorectal liver metastases (CRLM), neoadjuvant chemotherapy can downstage tumors, potentially making surgery feasible. We investigate whether radiomic signatures-quantitative imaging biomarkers derived from baseline computed tomography (CT) scans-can noninvasively predict chemotherapy response in patients with unresectable CRLM, offering a pathway toward personalized treatment planning. Approach We used radiomics combined with a stacking classifier (SC) to predict treatment outcome. Baseline CT imaging data from 355 patients with initially unresectable CRLM were analyzed using two regions of interest (ROIs) separately (all tumors in the liver and the largest tumor by volume). From each ROI, 107 radiomic features were extracted. The dataset was split into training and testing sets, and multiple machine learning models were trained and integrated via stacking to enhance prediction. Logistic regression coefficients were used to derive radiomic signatures. Results The SC achieved strong predictive performance, with an area under the receiver operating characteristic curve of up to 0.77 for response prediction. Logistic regression identified 12 and 7 predictive features for treatment response in all tumors and the largest tumor ROIs, respectively. Conclusion Our findings demonstrate that radiomic features from baseline CT scans can serve as robust, interpretable biomarkers for predicting chemotherapy response, offering insights to guide personalized treatment in unresectable CRLM.

Reliable and robust feature extraction is a fundamental step in machine learning pipelines that aim to extract additional information from medical images. Topological data analysis (TDA), a recent advancement based on the mathematical field of algebraic topology, has demonstrated significant success in various medical imaging domains. Persistent homology (PH) extracts complex topological features such as connected components, loops, and enclosed voids from data. For grid-structured data such as 3D computed tomography (CT) images, cubical complex is the standard method for PH computation; however, it does not achieve optimal performance. In this paper, we introduce a new approach to construct PH from volumetric medical imaging data. The backbone of this method is the transformation of images to a point cloud. This process utilizes principal component analysis (PCA) to compress patches of the volumetric data and converts to d-dimensional points; hence transforming the 3D image to a point cloud. Next, PH is constructed from the resulting point cloud using alpha complex filtration. Several experiments were conducted on two CT datasets benchmarked against the standard cubical complex method. Overall, the PCA-based approach outperforms the cubical complex algorithm in terms of classification performance on both datasets in all comparisons.

Pancreatic cancer is a uniformly deadly disease. Prediction of response to neoadjuvant therapy is critical in determining which patients should undergo invasive surgery. Non-invasive biomarkers of response would address gaps in the management of patients. This study employs pre-trained convolutional neural networks (CNNs) to predict response from baseline computed tomography (CT) scans alone, prior to neoadjuvant therapy. ResNet50, InceptionV3, VGG16, and Xception were trained on a dataset of annotated CT scans of patients with pancreatic ductal adenocarcinoma (PDAC). Our results demonstrate that the ResNet50 model achieves the highest performance among the models predicting response, with an average (average +/- margin of error at 95% confidence level) accuracy of 0.679 +/- 0.057, an F1-score of 0.665 +/- 0.072, recall of 0.717 +/- 0.081, precision of 0.698 +/- 0.074, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.781 +/- 0.162 across 5-fold cross-validation. These findings highlight the potential for non-invasive imaging biomarkers in predicting response to neoadjuvant therapy in PDAC.

Automatic identification of metastatic sites in cancer patients from electronic health records is a challenging yet crucial task with significant implications for diagnosis and treatment. In this study, we demonstrate how advancements in natural language processing, namely the instruction-following capability of recent large language models and extensive model pretraining, made it possible to automate metastases detection from radiology reports texts with a limited amount of gold-labeled data. Specifically, we prompt Llama3, an open-source instruction-tuned large language model, to generate synthetic training data to expand our limited labeled data and adapt BERT, a small pretrained language model, to the task. We further investigate three targeted data augmentation techniques which selectively expand the original training samples, leading to comparable or superior performance compared to vanilla data augmentation, in most cases, while being substantially more computationally efficient. In our experiments, data augmentation improved the average F1-score by 2.3, 3.5, and 3.9 points for lung, liver, and adrenal glands, the organs for which we had access to expert-annotated data. This observation suggests that Llama3, which has not been specifically tailored to this task or clinical data in general, can generate high-quality synthetic data through paraphrasing in the clinical context. We also compare metastasis identification accuracy between models utilizing institutionally standardized reports vs. non-structured reports, which complicate the extraction of relevant information, and show how including patient history with a customized model architecture narrows the gap between those two setups from 7.3 to 4.5 points on F1-score under LoRA tuning. Our work delivers a broadly applicable solution with remarkable performance that does not require model customization for each institution, making large-scale, low-cost spatio-temporal cancer progression pattern extraction possible.

The development of machine learning (ML) models based on computed tomography (CT) imaging has been a major focus due to the promise that imaging holds for diagnosis, staging, and prognostication. These models often rely on the extraction of hand-crafted features, incorporating robust feature engineering improves the performance of these models. Topological data analysis (TDA), based on the mathematical field of algebraic topology, focuses on data from a topological perspective, extracting deeper insight and higher dimensional structures. Persistent homology (PH), a fundamental tool in TDA, extracts topological features such as connected components, cycles, and voids. A popular approach to construct PH from 3D CT images is to utilize the 3D cubical complex filtration, a method adapted for grid-structured data. However, this approach is subject to poor performance and high computational cost with higher resolution CT images. This study introduces a novel patch-based PH construction approach tailored for volumetric CT imaging data that improves performance and reduces computational time. This study conducts a series of systematic experiments to comprehensively analyze the performance of the proposed method with various parameters and benchmarks against the 3D cubical complex algorithm and radiomic features. Our results highlight the dominance of the patch-based TDA approach in terms of both classification performance and computational time. The proposed approach outperformed the cubical complex method and radiomic features, achieving average improvement of 7.2

Colorectal cancer is the third most commonly diagnosed cancer in the United States with roughly half of these patients developing liver metastases, of which less than 10% survive past 3 years. Although resection is possible in some cases, only 15-25% of these patients are considered cured at the 10 year mark. Research supported by single-institution data shows that quantitative information taken from contrast-enhanced computed tomography (CECT) scans has the potential to pre-operatively predict patients at high risk for recurrence. CT image acquisition, reconstruction, and contrast timing affect the generalizability of predictive radiomic models in CT. Identification of reproducible features, therefore, is a necessary prerequisite to clinical implementation. We prospectively varied CT acquisition, reconstruction, and contrast timing parameters to quantify variability of radiomic features. Reproducibility analysis was performed using Lin's concordance correlation coefficient (CCC) with 135 patients from Memorial Sloan Kettering Cancer Center (n = 68) and University of Texas MD Anderson Cancer Center (n = 67). Each patient underwent an additional phase CECT scan within +/- 15 seconds of the routine portal venous phase using a controlled protocol, with systematic variations in scan timing, image acquisition, and image reconstruction. Radiomic features were extracted from the liver parenchyma and largest metastasis separately. Features extracted from the liver parenchyma were found to be less reproducible than those extracted from the largest tumor,. Reproducibility of features extracted from the tumor showed a negative correlation with the magnitude of scan delay. When reconstructed with a 5 mm slice thickness, scans with higher levels of adaptive statistical iterative reconstruction showed less reproducibility.

CT is a main modality for imaging liver diseases, valuable in detecting and localizing liver tumors. Traditional anomaly detection methods analyze reconstructed images to identify pathological structures. However, these methods may produce suboptimal results, overlooking subtle differences among various tissue types. To address this challenge, here we employ generative diffusion prior to inpaint the liver as the reference facilitating anomaly detection. Specifically, we use an adaptive threshold to extract a mask of abnormal regions, which are then inpainted using a diffusion prior to calculating an anomaly score based on the discrepancy between the original CT image and the inpainted counterpart. Our methodology has been tested on two liver CT datasets, demonstrating a significant improvement in detection accuracy, with a 7.9% boost in the area under the curve (AUC) compared to the state-of-the-art. This performance gain underscores the potential of our approach to refine the radiological assessment of liver diseases.

Computational competitions are the standard for benchmarking medical image analysis algorithms, but they typically use small curated test datasets acquired at a few centers, leaving a gap to the reality of diverse multicentric patient data. To this end, the Federated Tumor Segmentation (FeTS) Challenge represents the paradigm for real-world algorithmic performance evaluation. The FeTS challenge is a competition to benchmark (i) federated learning aggregation algorithms and (ii) state-of-the-art segmentation algorithms, across multiple international sites. Weight aggregation and client selection techniques were compared using a multicentric brain tumor dataset in realistic federated learning simulations, yielding benefits for adaptive weight aggregation, and efficiency gains through client sampling. Quantitative performance evaluation of state-of-the-art segmentation algorithms on data distributed internationally across 32 institutions yielded good generalization on average, albeit the worst-case performance revealed data-specific modes of failure. Similar multi-site setups can help validate the real-world utility of healthcare AI algorithms in the future.

Colorectal cancer is the third most common cancer globally, with a high mortality rate due to metastatic progression, particularly in the liver. Early diagnosis and effective treatment are critical for improving survival rates. Surgical resection remains the cornerstone for curative treatment, but only a small subset of patients are eligible for surgery due to the already advanced stage of the disease at the time of the diagnosis. For patients with initially unresectable colorectal liver metastases (CRLM), neoadjuvant chemotherapy can downstage tumors, potentially making surgical resection feasible. Predicting treatment response is crucial for optimizing treatment strategies, as it can guide personalized approaches and avoid exposing patients to the toxicity of chemotherapy. In this study we explored quantitative image features employing radiomics, which were then analyzed using machine learning algorithms to predict treatment outcome, offering a non-invasive tool to assist clinical decision-making and tailor personalized treatment strategies for CRLM patients. The study involved 355 patients with initially unresectable CRLM. Our findings demonstrate that baseline computed tomography scans contain valuable prognostic information. Radiomic-based models achieved high predictive performance, with the area under the receiver operating characteristic curve values reaching 77% for the a priori prediction of treatment response.

Assess the potential added benefit of radiomics to clinical models for predicting postoperative pancreatic fistula (POPF) after pancreatoduodenectomy (PD). Radiomics extracts quantitative data from medical imaging based on enhancement patterns. Clinical applications of radiomics have been investigated in pancreas, lung, breast, and prostate cancers. This single center retrospective study included all patients with available preoperative CT scan before PD from 2009-2021. Radiomic features that reflect heterogeneity in enhancement patterns were extracted from manually segmented future remnant pancreas. Clinical variables and radiomic features were used with random forest classifier to design five predictive models for grade B/C clinically relevant POPF (CR-POPF): preoperative (PreClin), intraoperative (IntraClin), radiomics alone (Rad), PreClin with radiomics (PreClin-Rad), and IntraClin with radiomics (IntraClin-Rad). Training data was randomly selected and comprised 70% (n=339) of the cohort, and the remaining 30% (n=145) was the test set. A prospective validation cohort (n=60) was also created. From 1855 eligible PD, all 234 with POPF were included, random sampling was used on the remainder to select 250 without POPF. In the test set, PreClin (AUC=0.74), IntraClin (AUC=0.78), and Rad (AUC=0.75) performed similarly. Best results were noted with combined models, AUC=0.82 for PreClin-Rad and AUC=0.84 for IntraClin-Rad. The same patterns were noted in the prospective validation cohort with PreClin-Rad performing best (AUC=0.78). Radiomics compared favorably with clinical risk scores for CR-POPF after PD, and combined models with radiomics and clinical data offer the strongest prediction. The PreClin-Rad model demonstrates the greatest clinical utility with excellent predictive outcomes relying entirely on preoperative data in the test and prospective validation cohorts.