Pancreatic cystic neoplasms are being found more and more often as abdominal CT and MRI scans become routine. These lesions include several types: serous cystic neoplasms (SCNs), mucinous cystic neoplasms (MCNs), and intraductal papillary mucinous neoplasms (IPMNs). Their malignancy potential differs dramatically, with SCNs being almost always benign, while IPMNs can become cancerous in up to 100% of cases.
The problem is that imaging alone often gets it wrong. Relying purely on radiological features misdiagnoses about one-third of cases, leading to either unnecessary surgeries with significant complication risks or delayed treatment of cancers that could have been caught earlier. Both MRI and CT achieve only 40-95% and 40-81% accuracy respectively for cyst type identification.
The researchers developed a Multimodal Deep Forest (MDF) model, trained on 449 MRI and 568 CT scans from patients diagnosed between 2009 and 2022. Crucially, the model integrates both imaging features (cyst size, location, mural nodules, pancreatic duct dilation) and clinical data (smoking history, diabetes, CA19-9 levels, and other lab results) rather than relying on images alone.
The Deep Forest approach uses cascaded tree ensembles, which automatically learn interactions between features and are particularly well-suited to medium-sized medical datasets where deep neural networks might overfit. To handle differences between CT and MRI data, missing modality-specific features were carefully coded rather than simply ignored, enabling the model to work robustly with patients who had only one type of scan.
The multimodal model achieved an average accuracy of 91.2%, precision of 91.7%, sensitivity of 88.9%, and specificity of 96.5% for classifying PCN types. These results significantly outperformed models using only imaging data or only clinical data, confirming that combining multiple data sources is essential.
Most remarkably, the model achieved 100% prediction accuracy for mucinous cystic neoplasms (MCNs) in patients imaged by either CT or MRI. Since MCNs have a malignancy rate of 10-39% and surgical resection is often recommended, perfect accuracy in identifying them would directly prevent unnecessary operations and ensure appropriate treatment.
When tested on an additional public pancreatic cancer dataset not used in training, the model again achieved top performance, demonstrating that it had learned generalizable patterns rather than memorizing its training data.
The 100% MCN detection rate has direct clinical significance: MCN patients can be confidently identified and referred for surgery without relying on subjective radiologist interpretation, while non-MCN patients can be spared unnecessary operations. This could substantially reduce the healthcare burden of over-treatment.
The model's high specificity of 96.5% means it rarely misidentifies benign cysts as dangerous, which is particularly important in clinical practice since false positives lead to anxiety, additional testing, and risky surgeries. For a disease where the treatment can be as dangerous as the condition, a highly specific AI tool provides significant safety value.
This study shows that integrating imaging data with clinical history through AI creates a diagnostic system superior to any single data source or human interpretation alone. The Multimodal Deep Forest framework requires minimal hyperparameter tuning, making it practical to deploy in clinical settings without specialized AI expertise.
The model's strong performance on both CT and MRI data, and its validation on independent datasets, positions it as a generalizable tool ready for broader clinical evaluation. Future work incorporating additional modalities such as cyst fluid biomarkers could push accuracy even higher.