Deep Learning in Endoscopic Ultrasound: Integrating Spatial Heterogeneity for Enhanced Diagnostic Precision in Distal Cholangiocarcinoma Detection

Aims Distal cholangiocarcinoma (dCCA) remains a diagnostic challenge due to its complex presentation and the limited specificity of traditional imaging methods. While previous artificial intelligence (AI) models have demonstrated high diagnostic accuracy, this study seeks to enhance the performance of convolutional neural network (CNN)-based models by incorporating spatial heterogeneity, a novel parameter that measures pixel distribution variance, to better capture the irregular structures characteristic of malignant tumors.

Methods Building upon previous findings, the dataset comprised 848 endoscopic ultrasound (EUS) images, generated from an initial 156 EUS images of patients diagnosed with dCCA cases through data augmentation. Using MATLAB’s DeepLabv3+model architecture, spatial heterogeneity was computed as the variation in pixel intensity within each detected tumor region. This parameter was integrated into the CNN for segmentation, with model accuracy, sensitivity, specificity, Intersection over Union (IoU), and Mean Boundary F1 Score used for performance evaluation. A comparative analysis was conducted using a UNet model implemented in Python [1].

Results Integrating spatial heterogeneity increased the classification accuracy to 98.5%, with sensitivity reaching 97.1% and specificity at 100%. Segmentation outcomes improved, with global accuracies of 87% for the pancreas and 92% for the bile duct., with a Mean IoU of 0.65 for the pancreas and 0.57 for the bile duct. The MATLAB DeepLabv3+model with spatial heterogeneity outperformed the UNet model in accurately delineating complex tumor boundaries, demonstrating improved performance in cases with high textural heterogeneity.

Conclusions Including spatial heterogeneity significantly enhances CNN-based EUS imaging models for dCCA, providing a refined approach for distinguishing malignant tissue with complex internal structures. This approach highlights the value of texture-based parameters in AI-driven diagnostics and supports the broader adoption of such techniques in clinical practice. Further research will focus on validating these findings in diverse patient cohorts.

Publikationsverlauf

Artikel online veröffentlicht:
27. März 2025

• References

• 1 Orzan R.I., Santa D., Lorenzovici N., Zareczky T.A., Pojoga C., Agoston R., Dulf E.-H., Seicean A.. Deep Learning in Endoscopic Ultrasound: A Breakthrough in Detecting Distal Cholangiocarcinoma. Cancers 2024; 16: 3792