by Christos Evangelou, MSc, PhD – Medical Writer and Editor
In a recent publication, researchers at the University of Shanghai for Science and Technology and the Naval Medical University in Shanghai, China, present a novel artificial intelligence (AI) model that demonstrates high accuracy in identifying various pancreatic abnormalities from CT scans.1
Though external validation is needed, this new lightweight, end-to-end multimodal network could significantly enhance early detection and treatment of pancreatic disorders, including pancreatic cancer, by combining rapid processing times with high accuracy.
The report was published in Clinical Radiology.
Study Rationale: Improving Pancreatic Disease Detection
Pancreatic diseases, ranging from inflammatory conditions to malignant tumors, pose significant challenges in diagnosis and treatment. Pancreatic cancer, in particular, has a very poor prognosis, with a 5-year survival rate of 7.9% to 33.9%.2 It is projected to become the second leading cause of cancer-related deaths in the United States by 2030. Although computed tomography (CT) plays a crucial role in the diagnosis of pancreatic diseases, the increasing use of high-resolution CT scans has led to a surge in the number of images radiologists must analyze, increasing the risk of missed diagnoses.1
The advent of deep learning in medical imaging has shown promise in addressing these challenges. However, previous studies have often focused solely on pancreatic cancer, neglecting other pancreatic disorders. In addition, many existing models require manual delineation of the pancreatic region, limiting their practical applicability in clinical settings.1
Methodology: A Novel Approach to Pancreatic Disease Detection
To address these limitations, researchers developed eeMulNet, a lightweight end-to-end multimodal network designed for rapid and accurate diagnosis of pancreatic abnormalities. The architecture of eeMulNet consists of two primary components: one for pancreatic region localization and one for multimodal CT diagnosis.1
The model employs a 3D nnUnet network to automatically segment the pancreas from CT scans, eliminating the need for manual region-of-interest delineation. Moreover, the model integrates both imaging data and patient information (age and gender) to allow clinicians to make a diagnosis using multimodal clinical data.
The image processing uses a three-dimensional adaptation of the MobileNetV2 network, which has been shown to have high efficiency and small parameter size. Patient information is processed using a Bert-Chinese pre-trained model. The extracted features and patient information are then fused using a Gated Recurrent Unit (GRU) before final classification.1
The researchers used a dataset of 943 CT scans, including 715 scans from patients with various pancreatic diseases and 228 from healthy individuals (controls). The dataset was split into training (755 scans) and independent testing (188 scans) sets. The model was trained using five-fold cross-validation to enhance generalizability.1
Model Performance
On the independent test dataset of 188 CT scans (143 abnormal, 45 normal), the eeMulNet model showed high performance in differentiating pancreatic abnormalities from normal pancreas, providing an area under the curve (AUC) of 1.0, accuracy of 100%, specificity of 100%, and sensitivity of 100%.1 These results indicate that the model correctly identified all cases of pancreatic abnormalities without any false positives or negatives. Moreover, eeMulNet took, on average, 41.04 seconds per patient to process CT images, with an average of 0.04 seconds required for classification.1
Importantly, the eeMulNet model outperformed several of the commonly used 3D convolutional networks, including various ResNet architectures.1 Its lightweight design resulted in a parameter size of only 3.8 megabytes, compared to 33–63 megabytes for the ResNet models, making it more suitable for real-world deployment.
Potential Implications
According to the authors, the high performance of eeMulNet in differentiating pancreatic abnormalities from normal pancreas demonstrates the model’s potential for reducing missed diagnoses and improving overall diagnostic accuracy. Moreover, the rapid processing time and automated detection of regions of interest could substantially streamline radiologists’ workflow, allowing them to focus on more complex cases.
Unlike previous models that focused primarily on pancreatic cancer, eeMulNet was trained on CT scans from patients with various pancreatic disorders. Training using this comprehensive dataset makes the model a more versatile tool for clinical practice. Furthermore, the lightweight design of the model makes eeMulNet suitable for deployment in various healthcare settings, including those with limited computational resources.
Limitations and Future Work
Despite the promising accuracy of eeMulNet, the performance of the model was evaluated on a dataset obtained at a single institution. External validation across multiple institutions is necessary to confirm the generalizability of the model. In addition, the researchers used 1 mm high-resolution portal phase CT scans, which may not be universally available. Future work should assess the model’s performance on various CT protocols and resolutions.
The authors also acknowledge that, in its current form, the model only distinguishes between normal and abnormal pancreatic cases. Future iterations could aim for a more granular classification of specific pancreatic diseases. Furthermore, the model currently only incorporates age and gender information. Inclusion of additional clinical data could further enhance its diagnostic capabilities.
In addition to conducting external validation studies across diverse healthcare settings and expanding the model’s capabilities to distinguish between different types of pancreatic diseases, future studies are needed to evaluate the model’s performance in real-world clinical settings. Moreover, future real-world studies are needed to assess the effect of eeMulNet implementation on diagnostic accuracy, workflow efficiency, and patient outcomes.
The study received financial support from the National Natural Science Foundation of China, the 234 Platform Discipline Consolidation Foundation Project of Changhai Hospital, and the Shanghai Science and Technology Innovation Action Plan Medical Innovation Research Project.
References
- Zhang G, Gao Q, Zhan Q, et al. Label-free differentiation of pancreatic pathologies from normal pancreas utilizing end-to-end three-dimensional multimodal networks on CT. Clin Radiol. Published online June 11, 2024. doi:10.1016/j.crad.2024.06.006
- Li J, Li Y, Chen C, Guo J, Qiao M, Lyu J. Recent estimates and predictions of 5-year survival rate in patients with pancreatic cancer: A model-based period analysis. Front Med (Lausanne). 2022;9:1049136. Published 2022 Dec 8. doi:10.3389/fmed.2022.1049136
