A deep learning architecture combining multi-phase computed tomography features from both the primary tumor and liver parenchyma accurately forecasts the development of early liver metastases in patients with resectable pancreatic ductal adenocarcinoma, according to a multi-center retrospective study published in Nature Communications. Pancreatic ductal adenocarcinoma remains one of the most lethal oncological malignancies, maintaining an overall five-year survival rate of approximately thirteen percent despite progressive developments in systemic therapies and diagnostic imaging protocols. Although curative-intent surgical resection provides the definitive cornerstone of therapeutic management for localized disease stages, postoperative recurrence affects up to eighty percent of patients, with the liver representing the most frequent anatomical site of distant metastatic dissemination. A substantial fraction of these hepatic recurrences develops within the initial six months following surgical resection, defining early liver metastases that reflect aggressive biological characteristics and preexisting occult micrometastatic disease. Patients presenting with early hepatic relapse typically experience dismal survival durations of less than six months post-recurrence, rendering upfront surgical extirpation minimally beneficial while imposing severe perioperative physiological stress and delaying therapeutic systemic intervention.
To address the critical clinical requirement for reliable preoperative risk stratification, the investigators assembled an extensive multi-institutional cohort encompassing one thousand and sixty-three patients across five medical centers in China. The development and retrospective evaluation cohorts comprised nine hundred and five patients who underwent upfront surgical resection, consisting of five hundred and thirty-eight males and three hundred and sixty-seven females with an overall mean age of 64.58 years, alongside an additional group of one hundred and fifty-eight patients who received neoadjuvant therapy prior to surgical resection, comprising seventy-eight males and eighty females with a mean age of 62.01 years. Patients undergoing upfront surgery were distributed into a primary training cohort of four hundred and fifty individuals, an internal validation cohort of one hundred and thirteen individuals derived from four contributing institutions, and an external validation cohort of three hundred and forty-two patients enrolled independently from Center Five. Across these distinct cohorts, primary tumors predominated within the pancreatic head or neck regions, and early liver metastases were observed postoperatively in 28.7 percent of the training cohort, 28.3 percent of the internal validation cohort, and 28.7 percent of the external validation cohort.
The deep learning framework was engineered around Mamba state space models, chosen for their linear recurrent formulation capable of preserving long-range global contextual dependencies across volumetric scans without incurring the prohibitive computational complexity typical of standard self-attention mechanisms. The computational framework incorporates a parallel dual-branch configuration wherein one specialized stream extracts features from the segmented primary pancreatic tumor and the complementary stream processes the surrounding liver parenchyma across non-contrast, arterial, and portal venous imaging phases. Because contrast dynamics differ markedly across temporal phases, abdominal scans were preprocessed using the nnU-Net framework, co-registering non-contrast and portal venous scans directly to the arterial phase, in which pancreatic lesions display optimal contrast margins. Initial volumetric shallow representations were extracted using three-dimensional depth-wise convolutions with a kernel size of 7 by 7 by 7, a padding of 3 by 3 by 3, and a stride of 2 by 2 by 2, before being fed into stacked PAN-Mamba encoders and progressive down-sampling blocks that simultaneously extract intra-phase volumetric spatial patterns and inter-phase temporal contrast variations.
Interaction between the separate organ features was established through an engineered Tumor-Organ Fusion module that integrates multi-phase representations via intra-organ self-attention and inter-organ cross-attention mechanisms. Within this fusion pipeline, concatenated multi-phase pancreatic and hepatic feature representations were refined by self-attention blocks with residual connections and layer normalization to eliminate redundant downsampling noise. Cross-attention layers then dynamically mapped interaction matrices between pancreatic tumor queries and liver key-value representations, as well as reciprocal projections between hepatic queries and pancreatic key-values. This architectural design allowed the algorithm to synthesize reciprocal microenvironmental signals between the primary lesion and its prospective metastatic niche before final classification. Computational implementation utilized PyTorch executed upon a single NVIDIA RTX 4090 graphics processing unit, training across one hundred epochs with binary cross-entropy loss, data augmentations including random noise and bias field distortions, and an initial learning rate of 1e-5 decreased systematically to 1e-7 through a cosine annealing schedule.
When assessed on classification efficacy, the Mamba-based model exhibited discriminatory precision in identifying early hepatic metastatic risk across independent validation datasets. In the training cohort of four hundred and fifty patients, the model achieved an area under the receiver operating characteristic curve of 0.890 with a 95 percent confidence interval of 0.860 to 0.920, demonstrating a sensitivity of 87.6 percent, a specificity of 78.8 percent, and an overall diagnostic accuracy of 0.813. The model sustained robust discriminative performance in the internal validation cohort, yielding an area under the curve of 0.820, sensitivity of 84.4 percent, and specificity of 72.6 percent. In the external validation cohort comprising three hundred and forty-two patients, the model maintained an area under the curve of 0.806 and an overall accuracy of 0.754. Comprehensive subgroup assessments confirmed stable discriminative capability across varied tumor locations, anatomical tumor sizes, preoperative serum carbohydrate antigen 19-9 concentrations, and diverse computed tomography scanner manufacturers from GE, Philips, Siemens, and Toshiba, while non-significant Hosmer-Lemeshow statistics verified strong calibration across testing sets.
Beyond binary metastatic classification at six months, the continuous risk probabilities generated by the Mamba model demonstrated pronounced prognostic utility for long-term patient survival outcomes. Patients were stratified into high-risk and low-risk categories utilizing an objective optimal cut-off value determined by the maximum Youden index derived within the primary training cohort. In a pooled univariable and multivariable Cox proportional hazards analysis of all nine hundred and five upfront surgery cases adjusting for conventional clinicopathological factors, model-assigned high-risk classification remained an independent predictor of adverse clinical outcomes. Specifically, high-risk status was significantly correlated with shortened progression-free survival, yielding an adjusted hazard ratio of 1.93 with a 95 percent confidence interval of 1.67 to 2.25, and shortened overall survival, yielding an adjusted hazard ratio of 1.89 with a 95 percent confidence interval of 1.61 to 2.23. Within the training cohort, high-risk individuals experienced a median progression-free survival of 4.8 months compared to 16.1 months in low-risk individuals, and a median overall survival of 13.5 months versus 33.3 months among low-risk patients.
The investigative team subsequently explored whether model-derived risk classifications could inform the clinical selection of patients for preoperative neoadjuvant systemic chemotherapy versus primary upfront resection. Evaluating three hundred and ninety-six patients treated at Center Five between June 2019 and December 2022, the researchers compared outcomes between one hundred and fifty-eight individuals who received neoadjuvant chemotherapy followed by surgical resection and two hundred and thirty-eight individuals who received immediate upfront surgery. To mitigate baseline selection imbalances between therapeutic pathways, a one-to-one propensity score matching analysis was implemented utilizing preoperative clinical variables, successfully isolating two balanced cohorts of one hundred and thirty-one matched patients each. Stratified survival comparisons revealed that high-risk patients who underwent neoadjuvant chemotherapy followed by resection achieved a median overall survival of 34.1 months compared to only 17.4 months for those receiving upfront surgery in the unmatched cohort, maintaining statistical significance following propensity score matching with a median overall survival of 34.1 months versus 18.2 months.
Conversely, among patients classified by the Mamba model as low risk for early hepatic recurrence, the administration of neoadjuvant therapy provided no statistically discernible overall survival advantage over upfront surgery either before or after propensity score matching. To rigorously address the confounding potential of immortal time bias and progressive disease occurring during the neoadjuvant administration window, the researchers executed a secondary landmark survival analysis calculating overall survival strictly from the date of completed surgical resection. This landmark assessment corroborated the primary findings, demonstrating a persistent survival extension exclusively in the high-risk neoadjuvant cohort. These therapeutic interactions suggest that imaging-derived occult metastatic risk profiles may offer objective biological criteria to refine patient selection for neoadjuvant chemotherapy regimens, helping clinicians avoid unproductive delays in low-risk operative candidates while selectively administering intensive preoperative systemic therapy to individuals harboring undetected micrometastatic disease.
To explore the molecular mechanisms captured by the noninvasive imaging signatures, the researchers conducted an exploratory radiotranscriptomic investigation in a subset of sixty-two patients from Center Five with matched preoperative imaging and tissue RNA sequencing. Total RNA extracted from formalin-fixed paraffin-embedded specimens was sequenced on the Illumina NovaSeq 6000 platform, and differential expression analysis contrasting model-defined high-risk and low-risk tumors identified fifty-four significantly downregulated genes meeting an adjusted false discovery rate threshold below 0.05 and absolute log2 fold-change exceeding 1.5. Kyoto Encyclopedia of Genes and Genomes pathway enrichment analyses revealed that these altered transcripts mapped to critical signaling cascades governing cellular proliferation, migration, and invasiveness, including the MAPK, PI3K-Akt, and Ras pathways. Furthermore, computational immune cell deconvolution via CIBERSORT demonstrated a distinct immunological microenvironment in high-risk tumors, characterized by a significant depletion of T follicular helper cells alongside a marked expansion of infiltrating monocytes, reflecting impaired antitumor immune surveillance and heightened biological aggressiveness.
Notwithstanding these robust results, the authors outlined key methodological limitations that contextualize the clinical interpretability of the current findings. The retrospective observational study design inherently carries potential residual confounding and patient selection biases, despite rigorous multivariable regression adjustments and propensity score matching. Furthermore, the study population was derived exclusively from East Asian tertiary medical institutions, and the observed baseline early liver metastasis incidence of approximately twenty-eight percent may not perfectly reflect Western surgical cohorts or broader global demographics. Additionally, the clinical utility analysis could not account for patients who initiated neoadjuvant chemotherapy but experienced interval disease progression that precluded subsequent curative-intent surgical intervention. Consequently, the authors conclude that prospective multi-center randomized validation across ethnically diverse geographic settings will be necessary before the Mamba-based radiomic signature can be fully integrated into routine clinical staging workflows or treatment guideline algorithms.
Subject of Research: Deep learning prediction of early liver metastases and neoadjuvant therapy response in pancreatic ductal adenocarcinoma using multi-phase computed tomography
Article Title: Deep learning CT signature for predicting early liver metastases in pancreatic ductal adenocarcinoma
Article References: Zhao, B., Gong, Z., Xiao, W., Chen, M., Huang, S., Zhu, L., Xia, T., Sun, Y., Yang, W., Liu, M., Ming, S., Yu, Y., Li, R., Xiao, Z., Xu, X., Chen, H., Song, Y., Tang, T., Lu, C., … Ju, S. (2026). Deep learning CT signature for predicting early liver metastases in pancreatic ductal adenocarcinoma. Nature Communications, 17(1), Article 10596. https://doi.org/10.1038/s41467-026-77665-z
Image Credits: AI Generated
DOI: 10.1038/s41467-026-77665-z
Keywords: Pancreatic ductal adenocarcinoma, Early liver metastases, Mamba architecture, Computed tomography, Neoadjuvant therapy, Overall survival, Progression-free survival, Radiotranscriptomics, Propensity score matching, Tumor microenvironment, Dual-branch deep learning, Surgical resection, Multi-center validation, CA19-9 biomarker, T follicular helper cells, Monocyte infiltration

