Early detection of pancreatic cancer on computed tomography: advancements with deep learning

Abstract

Advancements in artificial intelligence (AI) are transforming medical imaging diagnostics, offering new possibilities for automated pancreatic tumor detection in computed tomography scans. Pancreatic ductal adenocarcinoma continues to be one of the most lethal malignancies, with early detection being critical for improving survival rates. Deep learning models can learn hierarchical feature representations directly from imaging data, enhancing tumor detection accuracy. However, variations in model performance, impaired generalizability, and limited interpretability remain critical barriers to clinical adoption. This article provides a comprehensive overview of deep learning-based pancreatic tumor detection, discussing fundamental concepts, recent advancements, and challenges for clinical adoption. Implementation of deep learning tumor detection models into imaging workflows holds promise for improving early detection rates of pancreatic tumors. Addressing issues of standardization, external validation, and model transparency will be essential to enable the integration of AI into pancreatic cancer screening and diagnostics, ultimately improving early detection and patient outcomes.

Summary

This review explores the role of deep learning for pancreatic tumor detection in computed tomography, offering insights into advancements in detection models and challenges for clinical integration.

Essentials

• CT imaging remains the primary tool for pancreatic cancer diagnosis, yet conventional methods can miss nearly 40% of early-stage tumors.

• Deep learning models enhance pancreatic tumor detection by learning complex spatial patterns directly from imaging data, improving diagnostic accuracy.

• Challenges such as model generalizability, interpretability, and practical clinical integration must be addressed for widespread adoption.

• Efforts toward standardizing model development methods, evaluation, and reporting are key to improving reproducibility and reliability in real-world applications.

Introduction

In recent years, artificial intelligence (AI) has demonstrated its transformative power across different fields, with the greatest impact in data-intensive tasks that rely on pattern recognition such as medical imaging diagnostics.¹ Advances in computational power, coupled with refined AI architectures and access to large-scale training data, have driven the increased exploration of AI models for screening complex disease processes where early detection is critical for prognosis.² Among these, pancreatic ductal adenocarcinoma (PDAC) remains one of the most fatal malignancies, with a limited 5-year survival rate of just 13% across all disease stages despite advancements in treatment approaches.³ With the lack of specific early symptoms and rapid tumor progression, PDAC is often diagnosed at an advanced stage, limiting treatment options and significantly impacting survival outcomes. Contrast-enhanced computed tomography (CT) is the standard imaging modality for the diagnosis and initial evaluation of pancreatic cancer, offering high-resolution imaging of the pancreas and surrounding structures. Although early detection is crucial for improving prognosis,⁴ subtle early-stage tumors can often be overlooked due to their small size and resemblance to normal tissue. Conventional diagnostic methods still fail to identify nearly 40% of pancreatic tumors at an early stage,⁵ highlighting the need for improved detection strategies.

In the early phases of AI-driven tumor detection, 2 major approaches took the lead in the quantitative analysis of medical imaging features: radiomics-based machine learning models and deep learning (DL) approaches. Both techniques have shown promise in the detection of different pancreatic tumors.^6–12 While both extract quantitative features from medical images, they do so in fundamentally different ways. Radiomics models rely on predefined, handcrafted features derived from imaging data—such as shape, texture, and intensity-based metrics—requiring expert-driven feature selection and statistical analysis.¹³ In contrast, DL models automatically learn hierarchical features directly from the images, capturing complex patterns and learning connections without the need for manual feature engineering. The depth in DL models comes from their multiple layers of interconnected artificial neurons, which allow them to extract hierarchical features from the data, making them especially powerful for image analysis.¹⁴ This complexity also makes them more computationally demanding than radiomics-based machine learning models, as the increasing number of layers and parameters requires significant processing power.¹⁵ Even though both techniques might appear fundamentally different, growing evidence shows overlap and complementarity in their applications. Increasing implementations of DL have led to emerging hybrid approaches such as deep radiomics where DL complements or replaces parts of the radiomics pipeline. Leveraging the strengths of both hand-crafted radiomics features oriented to a target segmentation volume, and automatically learned deep features, can increase the performance and interpretation of models to facilitate their clinical implementation. Lesion detection models have already effectively used DL segmentation models to automatically delineate the radiomics extraction volume, bypassing the need for the labor-intensive manual segmentation.^6,16 Additionally, models integrating deep feature pipelines with layers of handcrafted radiomics are increasingly being tested to improve accuracy and explainability.¹⁷

As computational power and model architectures continue to evolve, DL has gained increasing attention for its ability to learn hierarchical feature representations directly from imaging data. DL models can leverage large-scale datasets to automatically capture complex spatial patterns, subtle tumor characteristics, and lesion boundaries, leading to enhanced robustness and diagnostic accuracy.¹⁸ Given the expanding body of evidence in DL-based tumor detection, a focused review separate from the radiomics-based approaches is warranted. This article provides a comprehensive overview of emerging DL methodologies, outlining fundamental concepts and practices in DL for medical imaging and discussing their potential to revolutionize early diagnosis in clinical practice. We offer insights into the current state of pancreatic tumor detection models and discuss their challenges for potential integration into clinical workflows.

Computer vision tasks for tumor detection

Computer vision is a branch of AI focused on enabling computers to analyze visual patterns, allowing them to “see” and interpret the content of an image. This visual understanding is achieved by modeling connections between extracted image features through neuron-like units with learnable synapses that represent the strength of feature interactions.¹⁹ In the context of early tumor detection, DL models leverage computer vision to perform key tasks in anomaly detection such as image classification and semantic segmentation across different image modalities.²⁰

One fundamental operation enabling DL-based algorithms to systematically analyze images is the convolution. A convolution applies trainable filters over sections of an image to extract meaningful features such as edges, textures, and shapes while preserving spatial relationships. Stacking multiple convolutional layers, along with other operational layers describing relationships between features, allows DL architectures to create a hierarchical understanding of visual data. An example of this is the U-Net, an encoding-decoding architecture commonly used for segmentation tasks to simultaneously capture broad contextual information and fine image details.²¹ The encoder progressively reduces spatial resolution to capture high-level features, followed by the decoder, which is applied to restore spatial resolution, enabling precise anatomical delineation of input images.

Image classification

Classification is a supervised learning task, meaning the user has control over the input data and the corresponding expected output during training. In this context, the goal is to train an algorithm to recognize patterns in labeled data to categorize the input (eg, CT scans) into predefined categories (eg, “normal pancreas” vs “PDAC”). DL classification models extract relevant features using operations such as convolutional layers, which learn to identify distinguishing patterns from labeled training examples. Convolutional neural networks (CNNs) leverage these learning mechanisms to recognize subtle variations in texture and density that may not be immediately visible to the human eye, and in the case of tumor detection, might indicate malignancies.²⁰ Once trained, these models can generalize learned features to new cases, enabling classification of unseen images (Figure 1). A key advantage of standalone classification DL models is their ability to focus on global features, allowing generalization from fewer training examples. Additionally, classification tasks typically require less intricate annotations, as images are labeled globally rather than requiring detailed region-specific markings. However, achieving a reliable classification still requires high-quality training labels. To ensure training accuracy in tumor detection, ground truth labels are often derived from histological analysis, providing a robust foundation for learning and validation.

Figure 1.

A histology-informed deep learning framework for CT image classification. A convolutional neural network (CNN) processes CT scan images to extract feature maps and predict an image label. The predicted label is compared to the histology-informed ground truth label, based on histopathological analysis. A loss function computes the error between the predicted and the true labels, enabling a backpropagation algorithm to modify the weights in the CNN model and minimize the prediction error. This process is performed iteratively until the calculated loss stabilizes, and the prediction accuracy no longer improves. Integration of histological labels aims to enhance the accuracy of the model in medical image classification.

Classification models and performance metrics

A reliable assessment of classification models requires splitting the data into training/validation sets for development, and reserving a separate testing set to evaluate generalization. This can determine the presence of overfitting, where models perform well on training data but fail on new unseen cases. Additionally, performing an external validation using independent datasets provides a more robust evaluation of performance and generalizability across diverse populations or acquisition conditions. Classification models for tumor detection are often assessed using several complementary key metrics. Accuracy measures the overall correctness of the classification (ie, distinguishing tumor vs no tumor cases); however, this measure alone can be misleading in highly imbalanced tasks, such as detecting low-incidence tumors. Sensitivity, or recall, evaluates the ability of a model to detect cases with tumors (true positive rate), while specificity quantifies its ability to correctly identify negative controls (true negative rate). Metrics like precision (positive predictive value) also play an important role in tumor detection, where false positives are costly and can lead to anxiety and unnecessary invasive procedures. Low-precision AI models can also result in reduced trust in AI-assisted diagnostics. The F1-score provides a balanced measure between the precision and recall of a model. Visual metrics, such as the receiver operating characteristic (ROC) curve, help illustrate the trade-off between sensitivity and the false positive rate, while the area under this curve (AUC-ROC) reflects the overall discriminative ability of a model. Together, these metrics provide a comprehensive performance evaluation of classification models, ensuring reliability for real-world applications.

Image segmentation

In CT images, the semantic segmentation task involves assigning voxel-level labels to delineate and isolate specific areas of interest (Figure 2). Unlike standard classification, which assigns a single label to an image, semantic segmentation provides a detailed, voxel-level understanding. DL models can automatically delineate tumors and organ structures from surrounding tissues, improving anatomical understanding, lesion localization, and tumor burden estimation. An advantage of segmentation architectures over classification models for lesion detection is their ability to precisely localize tumor. However, segmentation models usually require larger, high-quality training datasets due to the increased complexity of the task. Obtaining these datasets is especially challenging, as it involves a labor-intensive, expert-labeled manual annotation, with target structures meticulously labeled voxel by voxel.²²

Figure 2.

AI-based segmentation pipeline for medical image analysis. This pipeline displays a 3D U-Net model used to perform automated segmentation of CT scan volumes. The top section illustrates the U-Net architecture, an encoder-decoder architecture that processes a stack of CT slices using 3D convolutions (blue) to extract spatial features, followed by downsampling pooling layers (green) and upsampling layers (red) for spatial reconstruction, and a final softmax layer (yellow) to convert the output into voxel-wise probabilities. In the training phase, the segmentation network takes a CT scan volume as input to the encoder-decoder architecture to produce segmentation predictions. Model outputs are compared to expert-labeled reference segmentation masks, and discrepancies are used to iteratively readjust the weights through backpropagation, refining their accuracy over successive training iterations. During inference, a trained model processes a new CT scan, generating an AI-labeled segmentation mask.

Segmentation models and performance metrics

Similar to the classification task, segmentation models should be assessed separately on independent test sets. However, unlike classification, segmentation requires additional voxel-wise accuracy, making the evaluation more intricate. Metrics such as the Dice similarity coefficient (DSC) and the Normalized Surface Dice (NSD) have been specifically used for this task. The DSC measures the spatial overlap between predicted and ground truth masks, with higher values indicating better segmentation quality.^23,24 On the other hand, the NSD considers boundary-based discrepancies, making it particularly useful in cancer imaging where precise contour delineation and boundary determination are critical.²⁵ While overall detection accuracy is important, precise tumor localization relies on segmentation metrics to ensure reliability of model predictions. Nonetheless, ground truth annotations used as the reference standard typically rely on expert annotations that can vary between observers, leading to an expert-specific bias.²⁶ Because of this intrinsic variability, segmentation algorithms should focus on delivering reproducible results with strong generalization, rather than aiming to perfectly match a specific reference.

New architectures

Advancements in DL architectures have led to different configurations that improve performance across different tasks. Hybrid models, such as those integrating segmentation and classification within a single framework, have led to improved results in tumor detection, localization and differentiation.¹⁰ Another key innovation in modern architectures is the use of vision transformers (ViTs), which replace traditional convolutions with attention mechanisms.²⁷ The attention mechanism enables a model to simultaneously process different parts of an input and weigh them according to the context, allowing models to focus on the most relevant features. This context-aware system enhances spatial understanding and preserves global relationships, and it is the mechanism powering most of the popular large language models (LLMs) these days. Recent developments in multimodal models, such as CLIP (Contrastive Language-Image Pretraining), demonstrate the potential of integrating visual and language information within a shared framework, allowing models to understand visual information in the context of natural language descriptions.²⁸ These new emerging DL architectures offer the ability to capture multiple relationships in an image, further enhancing performance in complex visual tasks.

DL-based pancreatic tumor detection models

The augmentation of computing power, paired with improved model architectures and increasingly refined training datasets, has led to the increasing exploration of DL-based models for pancreatic tumor detection in recent years. Numerous studies have explored DL approaches to improve the diagnosis of pancreatic tumors in CT imaging.^7,10,28–38 These models leverage classification, segmentation, and hybrid architectures to distinguish pancreatic lesions from normal pancreatic tissue and surrounding structures. In this setting, DL-based models could act as an independent screener, flagging potential abnormalities for evaluation that might otherwise go unnoticed by radiologists. Models like PANDA (Cao et al.), one of the largest pancreatic lesion detection models for non-contrast CT, have shown the ability to enhance radiologist performance in challenging diagnostic scenarios. In a large real-world validation study, PANDA achieved an impressive 92.9% sensitivity for pancreatic lesion detection using CT without contrast enhancement, demonstrating its potential to improve pancreatic tumor detection in opportunistic non-contrast CT scan evaluations. Liu et al.³⁹ developed a segmentation-based pancreatic segmentation model that detects pancreatic anomalies and outputs the probability of these being a PDAC. With this model, they achieved the highest performance in the PANORAMA global challenge for PDAC detection on contrast enhanced CT, achieving an AUC-ROC for detection of 0.926 on the held-out testing set used for final evaluation. Several studies have contrasted the performance of DL models with expert radiologists, with some achieving comparable accuracy,^37,40 while others have displayed superior detection capabilities, particularly for subtle or early-stage tumors.^10,41  Figure 3 shows a pancreatic tumor case initially missed by radiologists but detected by our in-house trained U-Net. Among published studies, the U-Net architecture remains the primary backbone for segmentation task models, with recent modifications such as the incorporation of ViTs within the processing framework enhancing its performance.^10,32 A representative selection of published DL developments for pancreatic tumor detection is presented in Table 1.

Figure 3.

Pancreatic neuroendocrine tumor (PNET) initially missed on clinical evaluation and retrospectively detected by AI. Arterial-phase CT images of a patient with a pathologically confirmed small (<2 cm) PNET (yellow arrow) are shown in axial (A), coronal (B), and sagittal (C) planes. Insets show expert manual segmentations: tumor (red), pancreas (yellow), and pancreatic duct (blue). The lesion was initially overlooked (false negative) by two experienced radiologists in a retrospective reader study.⁶ The output of a PNET detection model trained in-house is displayed as a 3D rendering in (D), showing tumor and duct probability maps over the pancreatic parenchyma. The AI model identified the lesion with a maximum tumor probability of 0.99. The ground truth tumor annotation is shown for comparison in (E), overlaid on the same 3D render.

Table 1.

Representative peer-reviewed studies of deep learning models for pancreatic tumor detection.

Study	Tumor subtypes	Model publicly available	Imaging modality	Training/validation cohort	Independent testing cohort	Comparison to radiologist performance	Deep learning architecture	Summary metricsa
Classification-based models
Yan et al., 202430	Pancreatic tumor, NOS	No	NC-CT	N = 519	N = 304	No	MSMP + ResNet50 (CNN-based)	Precision: 0.682, Recall: 0.905, F1-score: 0.779
Korfiatis et al., 202338	PDAC	No	CE-CT (Portal Venous Phase)	N = 1080	Internal: N = 1238 External: N = 274 (TCIA-CPTAC, MSD, NIH PCT) PreDx: N = 233	No	Attention ResNet	External test set: AUC: 0.90 (95%CI: 0.803-0.891), SENS: 0.87, SPEC: 0.82. PreDx test set: AUC: 0.91 (95%CI: 0.86-0.94), Acc: 0.84, SENS: 0.75, SPEC: 0.90.
Viriyasaranon et al., 202336	PDAC	No	CE-CT (portal venous or pancreatic phase)	Training N = 3010 Internal validation N = 1277	External: N = 361 (MSD +TCIA cohorts)	No	ShuffleNet V2 (CNN-based)	External set: AUC: 0.61, Acc: 0.825 (95%CI: 0.783-0.861), SENS: 0.817, SPEC: 1.00.
							PVT (Transformer-based)	External set: AUC: 0.80, Acc: 0.878 (95%CI: 0.840-0.908), SENS: 0.865, SPEC: 1.00.
Liu et al., 202041	PDAC	No	CE-CT (portal venous phase)	N = 412	N = 363 (MSD +TCIA cohorts)	Compared to reports Local Test2: SENS: 0.917	VGG network (CNN based)	Local Test2: AUC: 0.999 (95%CI: 0.998-1.000), Acc: 0.989, SENS: 0.990, SPEC: 0.989. US test set: AUC: 0.920 (95%CI: 0.891-0.948), Acc: 0.832, SENS: 0.790, SPEC: 0.976.
Segmentation-based models
Liu et al., 202539	PDAC	Yes	CE-CT (portal venous phase)	PANORAMA challenge49 N = 2238	PANORAMA challenge49 N = 957	No	nnU-Net + ResU-Net	AUC: 0.926 (95%CI: 0.908-0.943). Lesion-level AP: 0.724 (95%CI: 0.671-0.776).
CLIP- Universal Model (Liu et al., 2024)28	Pancreatic tumor, NOS	Yes	CE-CT	N = 3410	N = 6173	No	SWIN UNETR	Pancreatic tumor detection: SENS: 0.939, SPEC: 0.912. Pancreatic tumor segmentation: DSC: 0.62, NSD 0.826.
Dong et al., 202429	Pancreatic tumor, NOS	No	CE-CT	N = 281 (MSD 2018)	External validation N = 153	No	AMFF-Net (Transformer-based)	Pancreas DSC: 0.821, pancreatic tumor DSC: 0.570
Park et al., 202340	Solid and cystic tumors	No	CE-CT (portal venous phase)	N = 852	Internal TS: N = 603 External TS: N = 589	Two radiologists Rad-1: TS-1 SENS: 0.917, SPEC: 0.932 TS-2 SENS : 0.947, SPEC: 0.956 Rad-2: TS-1 SENS : 0.955, SPEC: 0.943 TS-2 SENS : 0.956, SPEC: 0.962	nnU-Net	TS-1: AUC: 0.91 (95%CI: 0.89-0.94), SENS: 0.902, SPEC: 0.851. TS-2: AUC: 0.87 (95%CI: 0.84-0.89), SENS: 0.833, SPEC: 0.827.
PAPNet (Li et al., 2023)31	Pancreatic cysts (SCN, MCN)	No	CE-CT (pancreatic phase, 30-35 s)	N = 107	No	No	U-Net	Mean DSC: 0.845 ± 0.163
Abel et al., 202143	Non-PDAC cystic lesions	No	CE-CT (portal venous phase)	N = 221 patients, 543 cysts	No	3 radiologists, (visual metrics)	nnU-Net	Mean SENS: 0.788, with 0.48 FP per case.
Combined segmentation-classification models
PANDA (Cao et al., 2023)10	Multiple (PDAC, PNET, SPEN, IPMN, MCN, SCN, CP, OTHER)	No	NC-CT	N = 3208	N = 6239	Reader studies with multiple readers and configurations	U-Net segmentation + multi-task CNN classification + transformer classification	DSC whole pancreas: 0.903 (95%CI: 0.895-0.910). DSC PDAC: 0.719 (95%CI: 0.674-0.758) Overall pancreatic lesion detection: AUC: 0.984 (95%CI: 0.980-0.987)
Chen et al., 202337	PDAC	Yes	CE-CT (portal venous phase)	N = 1023	N = 1473	Compared to reports Local TS: SENS: 0.961	Segmentation UNet + classification CNN	Local TS: AUC: 0.96 (95%CI: 0.94-0.99), SEN: 0.899, SPEC: 0.959 Nationwide TS: AUC: 0.95 (95%CI: 0.94-0.96), SEN: 0.897, SPEC: 0.928
Alves et al. 202234	PDAC	No	CE-CT (portal venous phase)	N = 242	External: N = 361 (MSD +TCIA cohorts)	No	nnU-Net	External TS: Mean AUC: 0.889 (95%CI: 0.833-0.946)

^aSummary metrics from the highest performing model reported in the evaluation with less risk of bias (ie, cross-validation results, internal independent test set, OR external validation).

Acc = accuracy; AUC = area under the ROC curve; CE = contrast-enhanced; CI = confidence interval; CNN = convolutional neural network; CT = computed tomography; CV = cross-validation; DSC = Dice similarity coefficient; FP = false positives; MSD = The Medical Segmentation Decathlon⁴⁴; MSMP = multiscale and multiperceptron; N/A = not applicable; NC = non-contrast; NOS = not otherwise specified; PDAC = pancreatic ductal adenocarcinoma; PNET = pancreatic neuroendocrine tumors; PreDx = prediagnostic scans; SENS = sensitivity; specificity; SWIN UNET = Swin UNet Transformers; TCIA = The Cancer Imaging Archive⁴⁵; TS = test set; US = United States; ViT = vision transformer.

Small tumor detection

While any form of tumor detection can be valuable, models that perform well only on large, hard-to-miss tumors offer minimal benefit to the diagnostic workflow. Despite advancements in automated tumor diagnosis using DL models, detecting small pancreatic tumors in medical imaging remains one of the greatest challenges. Small lesions often exhibit only subtle visual cues that can be easily overlooked, even by experienced radiologists.⁵ Although the definition of a small tumor can be somewhat arbitrary, it is most commonly defined as one with the largest diameter of less than 2 centimeters (ie T1 stage). This threshold has demonstrated prognostic significance, with higher chances of tumor resectability and lower risk of metastasis at presentation. It is widely used in clinical guidelines and research studies to classify early-stage disease, ensuring consistency in diagnosis and treatment planning.⁴² Thus, a timely diagnosis of these small lesions is perhaps where AI models have the potential to make the greatest clinical impact.

Several studies already recognize the importance of differentially evaluating small tumor detection. An outline of the reported performance of published studies in small tumors is presented in Table 2. Alves et al.³⁴ aimed to train a model for small (<2 cm) tumor detection, achieving an AUC of 0.876 in the small PDACs of their independent test set. Chen et al.³⁷ conducted a stratified analysis of their model by tumor size, reporting a sensitivity of 86.4% in their local test set and 74.7% in a nationwide test set for PDACs under 2 cm compared to over 90% sensitivity for larger tumors. Viriyasaranon et al. explored the performance of different DL models using artificially generated pseudo-tumor (ie, computer simulated tumors inserted in the training set images) segmentations to improve training efficiency and reduce the segmentation burden. They found that the tumor detection accuracy of their trained models dropped significantly for T1 tumors, falling from 94.3% and 95.7% across all tumor stages to 54.0% and 55.3% in small tumors for their CNN and ViT models, respectively.³⁶ Park et al.,⁴⁰ in a multivariable sensitivity analysis of their U-Net DL model, found that the diagnostic performance was the lowest for subcentimeter cystic lesions, with sensitivity dropping to 38% and 16% in their 2 testing sets. Korfiatis et al.³⁸ went a step further by conducting an additional validation of their CNN-based classification model on a curated testing set of prediagnostic scans, achieving an 84% diagnostic accuracy. Testing on prediagnostic scans—defined by the authors as presumed negative scans performed between 3 and 36 months before the clinical diagnosis of PDAC—offers a unique opportunity to assess the added value of DL models to identify subtle disease-related patterns beyond human perception. This is particularly important because models trained solely on diagnostic scans are inherently biased toward visible lesions and may fail to capture the earliest manifestations of disease. Despite promising results, most DL models still face challenges in detecting small pancreatic tumors, with performance often declining as tumor size decreases. Advancing these models through improved training strategies, multi-modal imaging, and large-scale comprehensive datasets could enhance early detection across multiple pancreatic tumor subtypes, ultimately improving prognosis and treatment outcomes.

Table 2.

Performance metrics of deep learning studies on small pancreatic ductal adenocarcinomas.

Study	Training cohort tumor size	Small tumor definition	Number small tumors (testing)	Metrics on small tumors	Comparison with radiologist in small tumors
PANDA (Cao et al., 2023)10	(PDAC) median 3.0 cm (IQR: 2.3-4.4).	T1 tumor	External Test: N = 283	Proportion of T1 detected tumors on external test: 0.92	No
Korfiatis et al., 202338	N/A	PreDx Scans	N = 100	PreDx cohort: AUC: 0.91 (95%CI: 0.86-0.94), Acc: 0.84, SENS: 0.75, SPEC: 0.90	No
Viriyasaranon et al., 202336	Unknown	T1 tumor	N = 63	ShuffleNet model, Acc: 0.54 PVT model, Acc: 0.553	No
Park et al., 202340	Median: 3.1 cm (IQR: 0.5-12.7)	<1 cm	(Solid) TS-1: N = 1, TS-2: N = 1 (Cystic) TS-1: N = 13, TS-2: N = 19	TS-1 (Solid): SENS : 1.00, TS-2 (Solid): SENS: 1.00 TS-1 (Cystic): SENS : 0.38 TS-2 (Cystic): SENS: 0.16	Radiologist-1: TS-1 (Solid): SENS : 1.00, TS-2 (Solid): SENS: 1.00, TS-1 (Cystic): SENS: 0.62, TS-2 (Cystic): SENS: 0.84 Radiologist-2: TS-1 (Solid): SENS : 1.00, TS-2 (Solid): SENS: 1.00, TS-1 (Cystic): SENS: 0.85, TS-2 (Cystic): SENS: 0.89
Chen et al., 202337	Median 2.9 (IQR: 2.1-4.4)	<2 cm	Local TS: N = 22, Nationwide TS: N = 91	Local TS: SENS : 0.864 Nationwide TS: SENS: 0.747	Radiology report. Local TS: SENS: 0.955
Alves et al. 202234	Median: 2.8 cm (IQR: 2.3-3.7)	<2 cm	N = 73	AUC: 0.876	No
Liu et al., 202041	Median: 3.0 (IQR: 2.2-4.6)	<2 cm	N = 38	Combined local test set: SENS: 0.921	Radiology report. Combined local TS: SENS: 0.895

Acc = accuracy; AUC = area under the ROC curve; CI = confidence interval; N/A = not applicable; PDAC = pancreatic ductal adenocarcinoma; PreDx = prediagnostic scans; SENS = sensitivity; SPEC = specificity; TS = test set.

Performance across different tumor subtypes

Perhaps the higher incidence and aggressive nature of PDAC, compared to other pancreatic tumors, has led most published detection models to focus specifically on distinguishing PDAC from a “normal” pancreas. While highly relevant in clinical settings, this approach is not always realistic. If the ultimate goal of pancreatic lesion detection models is to develop an automated screening tool, the range of potential lesions it will encounter is much broader. Encountering benign conditions (eg, chronic pancreatitis, benign cystic lesions, intrapancreatic splenules) as well as various premalignant and malignant conditions other than PDAC (eg, intraductal papillary mucinous neoplasms, neuroendocrine tumors, metastases) can significantly impact the model performance. For instance, small benign lesions may be misclassified as malignant, increasing the false-positive rate (Figure 4). Similarly, small or inconspicuous tumors can be overlooked by a model trained on an unrepresentative dataset, reducing the detection accuracy for these tumors. Some published models already account for this broader spectrum of pancreatic lesions in their training data.^10,40 For instance, the PANDA model by Cao et al.¹⁰ was trained on 8 different pancreatic lesion categories and demonstrated an outstanding lesion detection performance on non-contrast CT. However, its sensitivity for neuroendocrine tumors was notably lower than for other lesions, which could be the result of the limited representation in the training cohort and absence of contrast enhancement. Park et al. trained their model on different solid and cystic pancreatic lesions, demonstrating performance variability between these 2 categories.^10,40 These findings highlight some of the challenges models face in generalizing to a broader range of lesion types and emphasize the need for more inclusive training strategies that encompass the full spectrum of pancreatic lesions.

Figure 4.

False positive abnormality detection. Arterial-phase CT images from a renal donor patient with no pancreatic pathology. A deep learning model trained in-house to detect pancreatic neuroendocrine tumors (PNETs) incorrectly classified this case as a PNET due to the presentation of the vascular structure highlighted (circle) in the axial (A), coronal (B), and sagittal (C) views. Model abnormality predictions are shown as an overlay in the CT planes (D, E, F). The 3D renderings show the model prediction in relation to the pancreatic parenchyma (G) and vasculature (H).

Challenges for clinical implementation

Impaired generalization

Despite significant advancements in DL models for pancreatic tumor detection, several challenges remain before widespread clinical implementation can be achieved. A major challenge models face in achieving clinical applicability is generalizing performance beyond the training dataset. For instance, Liu et al.⁴¹ observed a significant performance decline in their CNN classification model, trained exclusively on an Asian population, when evaluated on an external cohort from the United States. They attributed the performance drop to differences in population characteristics and scanning parameters, highlighting the variability in model robustness across diverse patient groups.

Most models struggle to handle cases with anomalous anatomical variations or rare conditions, which are often underrepresented in training datasets. To illustrate this, the case of a patient with widespread fatty replacement of the pancreas and a neuroendocrine tumor at the pancreatic head border is presented in Figure 5. The performance of 3 top-performing, publicly available segmentation models (Figure 5B–D)^28,32,46 is contrasted against the manual tumor and pancreatic segmentation performed by an experienced radiologist (Figure 5A). In this case, all 3 models failed to capture the entire pancreas due to fatty replacement, which is a relatively common anatomical variation. TotalSegmentator, trained for normal anatomical structures only, does not provide tumor predictions, but the model segmentation accuracy for the gland can be assessed (Figure 4B).⁴⁶ However, for models designed to predict tumor regions, this variation resulted in either failure to correctly identify the tumor location (CLIP-UniversalModel²⁸) or complete omission of its presence (Vista3D³²). Although this behavior is not uncommon, improvements in DL architectures have led to better generalization of models to unseen cases and continuous enhancement of model predictions. Nonetheless, a widespread clinical application of DL models for automated pancreatic tumor screening will require proof of better generalization and robustness across different clinical scenarios.

Figure 5.

Performance of publicly available pancreas segmentation models in a real-world case of a patient with fat infiltration of the pancreas and a pathologically proven neuroendocrine tumor. (A) Manual segmentation: expert-labeled ground truth segmentation of the pancreas (yellow) and pancreatic tumor (red), with corresponding volumetric measurements. (B) TotalSegmentator⁴⁶: automated segmentation of the pancreas (green), contrasted against expert-labeled ground truth (DSC: 0.55). This model is not trained to provide tumor predictions. (C) CLIP-UniversalModel²⁸: automated segmentation of the pancreas (green), contrasted against expert-labeled ground truth (DSC: 0.68). Although a pancreatic tumor was predicted (purple), the location was incorrect with no overlap with the pancreatic tumor region (DSC: 0.0). (D) Vista3D: automated segmentation of the pancreas (green), contrasted against expert-labeled ground truth (DSC: 0.51). The case was a false negative with no pancreatic tumor detected. Moreover, an adrenal adenoma was incorrectly predicted as part of the pancreas parenchyma (inset). cc = cubic centimeters; DSC = Dice similarity coefficient.

Reproducibility challenges

Beyond generalization across data heterogeneity, variability in model development and reporting further hinders reproducibility of DL detection models. Differences in training methodologies, dataset preprocessing, ground truth definitions, and evaluation metrics can lead to inconsistent performance across studies, making it difficult to compare models or replicate results in different clinical settings. Without proper validation and standardized reporting, model predictions can vary significantly, raising concerns about reliability, fairness, and real-world applicability. Additionally, inconsistencies in annotation protocols and the lack of standardized ground truth definitions contribute to labeling discrepancies, further limiting reproducibility. Some models also lack external validation of their findings, preventing a clear assessment of their robustness when applied to new patient populations or imaging conditions.

To address these challenges, standardization initiatives such as CLAIM (Checklist for Artificial Intelligence in Medical Imaging) aim to establish guidelines for model development, validation, and reporting, ensuring greater consistency across AI models.^47,48 By harmonizing evaluation protocols and promoting transparent reporting, these initiatives help bridge the gap between research and clinical implementation of DL-based tumor detection models. Other efforts, such as openly sharing code and final model versions alongside published results, further enhance model evaluation, reproducibility, and continuous model improvement. Multi-institutional benchmarks, such as the recent PANORAMA challenge for PDAC detection,⁴⁹ also play a key role by providing standardized heterogeneous datasets and evaluation metrics that encourage transparent comparisons across algorithms and drive the progress of AI models. Despite these efforts, resulting models can still underperform in unfamiliar conditions, as shown in Figure 6 where impaired calibration led the model to assign a very low probability of PDAC despite correctly localizing the tumor and identifying ductal dilation. Achieving true reproducibility will require ongoing multi-institutional validation, homogeneous annotation strategies, prospective evaluations of performance, and potentially regulatory oversight, to ensure AI models perform reliably and consistently across diverse clinical settings.

Figure 6.

Multilabel model prediction with accurate tumor and duct localization, but impaired calibration. Venous-phase CT images from a patient with a confirmed pancreatic ductal adenocarcinoma (PDAC), visible in the axial (A), coronal (B), and sagittal (C) planes (yellow arrows). Insets show ground-truth segmentations for the tumor (red), pancreas (yellow), and duct (blue). Panels (D) and (E) display 3D renderings of the prediction from the winning model from the PANORAMA challenge.³⁹ Shown are anterior(D) and posterior (E) views, respectively, highlighting the model prediction for tumor regions (orange) and pancreatic duct (blue). Although the tumor and dilated duct were accurately identified, model calibration failed to classify this case as a PDAC with a maximum probability of 5.1%. Panel (F) shows the 3D rendering of the expert-labeled ground truth segmentation for the PDAC (red) and pancreatic duct (blue).

Addressing the “black-box” effect

An additional challenge specific to DL models in a clinical context is the “black box effect,” which refers to the lack of transparency in the way models make their final predictions. Unlike traditional diagnostic methods, where reasoning can be traced back to biological mechanisms, the complex learning process of neural networks and the hidden feature connections driving predictions make interpretability far more challenging. This lack of explainability, combined with the risk of model “hallucinations” and false positives, often raises concerns among clinicians, as understanding the rationale behind DL-driven diagnoses is critical for trust and clinical decision-making.

To address this challenge, ongoing efforts in Explainable AI focus on techniques such as saliency maps, feature attribution and attention mechanisms, which help visualize the most influential regions within an image that drive the prediction.⁵⁰ By enhancing transparency, these methods improve interpretability and foster clinician trust in DL-based diagnostic models.⁵¹ Integrating explainability frameworks, improving cross-cohort model validation, and aligning AI models with standardized reporting guidelines will be essential to ensure reproducibility, robustness, and move toward the widespread clinical adoption of automated pancreatic tumor screening.

Future directions

As DL architectures evolve, meaningful improvements in pancreatic tumor detection models will also require parallel advancements in training datasets. The advent of photon-counting CT provides a promising avenue to explore by offering higher spatial resolution and improved tissue contrast.⁵² At the same time, multimodal integration of different inputs, such as different image modalities, clinical history, laboratory values, and genomic information, can potentially build more comprehensive AI models that better reflect the complexity of real-world diagnosis and provide more robust AI screening and diagnostic tools.^53,54 Additionally, exploration of datasets specifically designed for AI feature extraction, as opposed to human visual interpretation, carries the potential of improving model understanding beyond human perception. The adoption of federated learning strategies can also play a critical role in the improvement of these detection models by enabling institutions to collaboratively train models on diverse datasets.⁵⁵

Leveraging electronic health records for early detection

One promising direction in multimodal detection is the analysis of longitudinal electronic health records (EHR). Emerging applications of LLMs show strong potential to extract information from unstructured clinical narratives and temporal trends, generating representations that can be interpreted by AI systems.⁵⁶ These models can potentially identify subtle and evolving signals, such as symptom patterns, laboratory value changes, or medication adjustments, that may appear before anatomical abnormalities become detectable. For instance, in the setting of pancreatic cancer, the onset of new or worsening diabetes has been recognized as a potential early marker of disease.⁵⁷ AI models may be able to detect concerning patterns in advance and integrate them with imaging and molecular data into multimodal prediction frameworks, enabling earlier risk stratification and more effective diagnostic and monitoring strategies.

Integration of systemic imaging biomarkers

One key advantage of DL models is their ability to evaluate medical images holistically and detect subtle cues in imaging biomarkers that may not directly involve the tumor itself. For instance, changes in body composition, such as reductions in skeletal muscle mass or alterations in fat distribution, can serve as early indicators of pancreatic disease and may help stratify patients who are at increased risk of malignancy.⁵⁸ Sarcopenia, for instance, has been associated with both onset and poor prognosis in pancreatic cancer. Depending on the diagnostic criteria, it can be identified in up to 65-70% of patients at the time of diagnosis.^59,60 DL models are especially well suited to recognize these nuanced features, analyzing not only the volume of skeletal muscle but also its quality, which may be overlooked by human observers in routine clinical practice.⁶¹ By incorporating such non-tumor-specific biomarkers, DL approaches could enhance predictive performance and contribute to both earlier detection and more accurate disease prognostication.

Combining liquid biopsy with computer vision

Another promising area of research is the integration of CT detection models with liquid biopsy techniques. Liquid biopsy provides a minimally invasive approach to detect tumor-derived material in blood or other body fluids, including circulating tumor DNA (ctDNA), circulating tumor cells, and exosomal markers. Although more commonly used for disease monitoring, these biomarkers have also demonstrated potential for identifying early molecular changes related to pancreatic cancer, often before abnormalities can be seen on imaging studies.⁴ Among these, ctDNA has garnered significant attention due to its ability to detect clonal somatic mutations, such as the KRAS mutations commonly found in pancreatic tumors.⁶² Studies have shown that combining ctDNA analysis with protein biomarkers (eg, CA19-9, CEA, HGF, and osteopontin) can significantly improve diagnostic sensitivity for pancreatic cancer.⁶³ However, ctDNA detection still displays low sensitivity in early-stage disease where ctDNA levels in blood are extremely low, restricting its use as an isolated screening tool.⁶⁴ Multimodal AI has the potential to address these challenges by combining liquid biopsy data with information from imaging studies, clinical history, laboratory results, and genetic risk factors. By analyzing large multimodal datasets, these models could uncover complex patterns that may not be apparent through a single input. For example, early molecular alterations identified in ctDNA can be interpreted in the context of subtle imaging changes or evolving laboratory trends, with dynamic thresholds to call anatomical alterations suspicious for malignant disease.

Conclusions

DL has demonstrated significant potential in automating pancreatic tumor detection in computed tomography, offering promising advantages for early diagnosis. By learning complex spatial patterns and subtle imaging biomarkers, these models can enhance detection accuracy and aid in the identification of small, inconspicuous tumors, complementing radiologists in the diagnostic workflows. However, several challenges must be addressed before clinical integration, including overcoming limited generalization, reproducibility issues, and the “black box effect,” which hampers interpretability and physician trust.

Ongoing advancements in AI architectures, coupled with the expansion of heterogeneous, multi-institutional datasets, will be critical for enhancing model performance and robustness. Bridging the gap between proof-of-concept studies and clinical application remains a critical challenge. Prospective validation of AI predictions in real-world clinical decision-making settings is essential to further elucidate the role of AI in early pancreatic cancer detection and assess its impact on patient outcomes. The future of early pancreatic cancer detection will depend on developing multimodal AI systems that are trained on rich and diverse sources of information, encompassing imaging, clinical records, molecular profiling, and systemic biomarkers. As these technologies evolve, rigorous validation, adherence to regulatory standards, and seamless integration into existing radiology workflows will be necessary to build trust among physicians and fully realize the potential of AI-driven pancreatic tumor detection in clinical practice.

Glossary

Abbreviations

AUC =

area under the ROC curve

CLIP =

Contrastive Language-Image Pretraining

CNN =

convolutional neural network

DL =

deep learning

DSC =

Dice similarity coefficient

LLMs =

large language model

NSD

= Normalized Surface Dice

PDAC =

pancreatic ductal adenocarcinoma

ROC =

receiver operating characteristic

Author contributions

Felipe Lopez-Ramirez (Conceptualization, Writing—original draft, Writing—review & editing), Emir A. Syailendra (Writing—original draft, Writing—review & editing), Florent Tixier (Writing—original draft, Writing—review & editing), Satomi Kawamoto (Writing—review & editing), Elliot K. Fishman (Conceptualization, Writing—review & editing), and Linda C. Chu (Conceptualization, Writing—review & editing)

Supplementary material

Supplementary material is available at Radiology Advances online.

Funding

This study was supported by a grant from the Lustgarten Foundation for Pancreatic Cancer Research.

Conflicts of interest

Felipe Lopez-Ramirez, Florent Tixier, Satomi Kawamoto, and Linda C. Chu receive salary support from The Lustgarten Foundation. Emir Syailendra receives salary support from the IPMN Global Foundation. Elliot K. Fishman reports grant support from the Lustgarten Foundation, Siemens, and is the co-founder of HipGraphics.