Multi-window scanning method for region-of-interest selection in pancreatic endoscopic ultrasound images.

This study presents a method for semantic segmentation of pancreatic endoscopic ultrasound (EUS) images. We developed a hybrid method to automatically select relevant frames from EUS video sequences, employing two nested dynamic windows: a smaller window scans local regions of each frame, while a larger window defines the overall scanning area, capturing spatial context. Texture features, including mode, mean, and standard deviation, are extracted from both windows to construct a feature vector for a fully connected neural network classifier (NNC), which predicts the class of each pixel. Following pixel-wise classification, a heatmap is generated for each frame to highlight regions of interest (ROI), allowing specialists to identify echotextural features. The method uses hierarchical image decomposition to improve ROI identification. Custom software was implemented to enable interactive segmentation, division into local windows, ROI classification, and structured database formation. Based on expert endoscopist evaluation, the optimal window size was determined to be 32 × 32 pixels. A curated dataset of local windows containing normal and pathological pancreatic echotextures was compiled from the selected frames. Experimental evaluation on 114 test frames demonstrated that the proposed method achieves an overall accuracy of 95.6% (PPV 96.3%; sensitivity 97.5%; specificity 91.2%), distinguishing relevant frames for subsequent semantic segmentation. This framework supports clinical assessment and automated ROI extraction in pancreatic EUS imaging.