Pancreatic neuroendocrine tumors (panNETs) are a distinct group of pancreatic cancers arising from hormone-producing islet cells rather than ductal epithelium. While generally less aggressive than pancreatic ductal adenocarcinoma, panNETs still carry significant mortality risk because approximately 60 to 90% of patients develop liver metastasis, which dramatically worsens prognosis.
The Ki67 proliferation index is the primary grading tool for panNETs, dividing tumors into grade 1 (Ki67 under 3%), grade 2 (Ki67 3 to 20%), and grade 3 (Ki67 above 20%). However, Ki67 alone fails to capture the spatial distribution of cell proliferation within the tumor, which may contain regions of varying aggressiveness.
Intratumoral heterogeneity in Ki67 expression is increasingly recognized as a key driver of treatment resistance and metastatic potential, but conventional pathology scores a single average value that masks this spatial complexity. A tumor with focal high-Ki67 regions surrounded by low-Ki67 areas may behave very differently from a homogeneously moderate-Ki67 tumor with the same average score.
There is an unmet clinical need for better preoperative predictors of liver metastasis risk in panNETs. More accurate risk stratification would allow surgeons and oncologists to tailor surveillance intensity, select patients for aggressive resection, and identify those who may benefit from systemic therapy before or after surgery.
The study introduced the Morisita-Horn (MH) index, an ecological diversity metric, to quantify spatial heterogeneity in Ki67 expression across panNET tissue sections. Originally developed to measure species distribution overlap in ecosystems, the MH index here measures how evenly or unevenly Ki67-positive cells are distributed across the tumor.
Whole-slide pathology images were divided into systematic grid regions. For each region, the local Ki67 positivity rate was calculated. The MH index then captures whether these regional rates are consistent across the tumor (low heterogeneity, high MH) or highly variable from region to region (high heterogeneity, low MH).
A Pathomics score was constructed by combining the mean Ki67 index with the MH heterogeneity index. This two-component score attempts to capture both the overall proliferation burden and the spatial architectural complexity of tumor growth, which together better reflect metastatic potential than either measure alone.
The Pathomics score was trained and validated to predict liver metastasis risk, achieving an area under the curve (AUC) of 0.799 in the validation cohort. While this represented a meaningful improvement over Ki67 index alone, researchers sought to further enhance predictive accuracy by incorporating radiological imaging features.
To complement pathomics, the study developed a deep learning radiomics (DLR) score from preoperative CT imaging. Radiomic features capture quantitative information about tumor texture, shape, and density that reflects underlying tissue biology not visible to the naked eye.
Deep learning feature extraction used the ResNET101 convolutional neural network, a 101-layer architecture pretrained on ImageNet, to extract high-dimensional texture representations from CT images of the primary tumor. These deep features were then combined with conventional hand-crafted radiomic features including shape descriptors, intensity statistics, and texture matrices.
Feature selection used LASSO (Least Absolute Shrinkage and Selection Operator) regression to identify the most informative subset of features while penalizing model complexity. The final DLR score was computed as a weighted linear combination of selected features, achieving an AUC of 0.875 in the validation cohort, outperforming the Pathomics score alone.
Importantly, the DLR score reflects preoperative imaging characteristics and therefore provides predictive information before surgery, making it potentially useful for treatment planning. By contrast, the Pathomics score requires surgical resection to obtain tissue, so it provides post-operative rather than pre-operative risk information.
The most powerful model combined the Pathomics score, DLR score, and independent clinical predictors into a single integrated nomogram. Among all clinical variables tested, nerve infiltration was the only independent predictor of liver metastasis in multivariate analysis, underscoring the aggressive biological behavior associated with perineural invasion in panNETs.
The integrated nomogram achieved an AUC of 0.985 in the training cohort and 0.961 in the validation cohort, representing near-perfect discriminative ability. This far exceeded the individual contributions of Ki67 alone, Pathomics score alone, or DLR score alone, demonstrating the synergistic value of combining pathological, radiological, and clinical information.
Calibration curves confirmed that the nomogram's predicted probabilities were well-aligned with actual observed liver metastasis rates across the full range of predicted risk, validating that the tool is reliable rather than simply discriminating between extreme cases.
Recurrence-free survival analysis showed that patients classified as high-risk by the nomogram had a median recurrence-free survival of 28.5 months compared to 34.7 months for low-risk patients, and the difference in long-term outcomes was statistically and clinically significant.
The integrated nomogram offers a practical tool for preoperative and postoperative decision-making in panNET management. Patients identified as high-risk for liver metastasis could be prioritized for more frequent postoperative surveillance imaging, earlier initiation of systemic therapy, or consideration for liver-directed interventions.
The identification of nerve infiltration as the sole independent clinical predictor of liver metastasis highlights an underappreciated biological mechanism. Perineural invasion likely provides tumor cells with a route for systemic dissemination, and its assessment should become a standard element of panNET pathological reporting.
The dual imaging-pathology design of the nomogram addresses a practical clinical need. Surgeons can use the DLR score from preoperative CT to counsel patients before surgery, while the full nomogram with Pathomics score provides refined risk stratification after resection, enabling dynamic, stage-appropriate risk communication.
Limitations include the retrospective single-center design involving 163 patients. Prospective multicenter validation in larger panNET cohorts will be needed before the nomogram can be recommended for routine clinical use, and standardization of CT acquisition protocols will be important for ensuring consistent DLR score calculation across centers.
This study demonstrates the power of multimodal integration in cancer risk prediction by combining three distinct data streams: spatial pathomics from whole-slide tissue analysis, deep learning features from CT imaging, and clinical pathological findings. Each layer captures a different dimension of tumor biology.
The Morisita-Horn index represents an innovative application of ecological diversity mathematics to cancer biology, providing a mathematically rigorous quantification of intratumoral heterogeneity that outperforms simple average Ki67 values for risk stratification.
The use of ResNET101 deep features from preoperative CT images allows the nomogram to extract predictive information from imaging that would otherwise require pathological analysis. This represents a step toward virtual biopsy, where imaging alone could provide tissue-level biological insights without requiring tissue sampling.
Future work should explore whether the nomogram can be updated dynamically with intraoperative or post-treatment data, and whether similar heterogeneity-based pathomics approaches could be applied to other gastrointestinal neuroendocrine tumors or to the prediction of responses to somatostatin analogs and targeted therapies in panNETs.