Aim: This study aimed to develop and validate a deep learning radiomics nomogram (DLRN) derived from ultrasound images to improve predictive accuracy for lymph node metastasis (LNM) in pancreatic cancer (PC) patients.
Methods: A retrospective analysis of 249 histopathologically confirmed PC cases, including 78 with LNM, was conducted, with an 8:2 division into training and testing cohorts. Eight transfer learning models and a baseline logistic regression model incorporating handcrafted radiomic and clinicopathological features were developed to evaluate predictive performance. Diagnostic effectiveness was assessed for junior and senior ultrasound physicians, both with and without DLRN assistance.
Results: InceptionV3 showed the highest performance among DL models (AUC = 0.844), while the DLRN model, integrating deep learning and radiomic features, demonstrated superior accuracy (AUC = 0.909), robust calibration, and significant clinical utility per decision curve analysis. DLRN assistance notably enhanced diagnostic performance, with AUC improvements of 0.238 (p = 0.006) for junior and 0.152 (p = 0.085) for senior physicians.
Conclusion: The ultrasound-based DLRN model exhibits strong predictive capability for LNM in PC, offering a valuable decision-support tool that bolsters diagnostic accuracy, especially among less experienced clinicians, thereby supporting more tailored therapeutic strategies for PC patients.
Keywords: Deep learning radiomics nomogram; lymph node metastasis; machine learning; pancreatic cancer; ultrasound.
Pancreatic cancer is a serious illness, and one of the most important factors in determining a patient’s outlook is whether the cancer has spread to nearby lymph nodes. Lymph nodes are small, bean-shaped structures that are part of the immune system and can act as pathways for cancer to spread throughout the body. However, it is often difficult to accurately detect this spread before surgery.In this study, we aimed to improve how doctors predict lymph node involvement by using detailed information from ultrasound images. We analyzed data from 249 patients who had been diagnosed with pancreatic cancer. By integrating measurable features extracted from medical images with relevant clinical data, we developed a predictive model known as a radiomics nomogram – an individualized risk assessment tool that estimates the likelihood of disease spread with improved accuracy.We found that this tool provided highly accurate predictions. It also helped doctors make better diagnostic decisions, especially those with less experience. When using the tool, both junior and senior ultrasound doctors were more likely to correctly identify patients with lymph node involvement.This approach has the potential to support doctors in making more informed treatment plans for pancreatic cancer patients and may help guide decisions before surgery.