Pancreatic cancer is one of the most deadly cancers, with fewer than 1 in 10 patients surviving five years. One major reason the prognosis is so poor is that cancer cells spread to nearby lymph nodes very early, often before a patient is even diagnosed.
Knowing whether lymph nodes are involved before surgery is critical — it determines the treatment plan and predicts how long a patient may survive. However, current imaging methods like CT and MRI are not reliable enough to detect lymph node spread beforehand.
Radiologists currently rely on visual inspection of scans, which is subjective and time-consuming. A more accurate, automated approach is urgently needed to improve treatment planning.
Researchers developed a deep-learning model called MTCN (Multiview-guided Two-stream Convolution Network) that analyzes CT scan images of both the tumor and the tissue surrounding it. The model was trained on 363 pancreatic cancer patients from a major hospital in Shanghai.
The network used two 'streams' of image analysis — one focused on the tumor itself, another on the surrounding tissue — to capture as much relevant information as possible. This was then combined with clinical factors like patient age and the blood marker CA125.
The final model, called MTCN+, merged AI-extracted image features with physician judgment and blood test results, creating a hybrid approach designed to outperform either alone.
The MTCN+ model achieved an accuracy of 76.1% in the training group and 76.1% in the test group, compared to just 63.3% for radiologist-only assessment. The area under the curve (AUC) — a measure of diagnostic ability — reached 0.815 in the test group versus 0.640 for radiologists alone.
In an external validation group from two additional hospitals, the MTCN+ model's AUC reached 0.854, confirming strong performance on patients it had never seen before. Roughly 40% of patients who were misclassified by radiologists could be correctly diagnosed using the AI model.
The model also performed well for patients with smaller tumors, which are notoriously harder to assess. Survival curves predicted by the model closely matched actual patient outcomes, showing the model has real prognostic value.
Knowing lymph node status before surgery allows doctors to decide whether a patient should receive chemotherapy first to shrink the cancer before operating. This approach, called neoadjuvant chemotherapy, can significantly improve survival in patients with lymph node involvement.
By replacing or supplementing subjective radiologist assessment with the AI model, hospitals could provide more consistent, objective pre-surgical evaluations. Patients who might otherwise be undertreated or overtreated would benefit from a more accurate picture of their disease.
The model is also promising for predicting overall survival, which could help patients and families have more informed conversations about prognosis and treatment goals.
The MTCN+ model represents a meaningful advancement in preoperative pancreatic cancer assessment. By combining deep learning image analysis with clinical data, it offers more accurate and objective lymph node status prediction than either approach alone.
One limitation is that the model was less effective at predicting the number of affected lymph nodes — only whether spread had occurred. Future work will aim to refine this capability, especially for patients already known to have lymph node involvement.
Overall, this research shows that integrating AI into routine CT scan analysis has real potential to improve surgical decision-making and patient outcomes in pancreatic cancer.