Pancreas Region Object Detection Based on Faster R-CNN

To address the challenges of low pancreas voxel ratio, complex background interference, and low contrast with surrounding organs in abdominal CT images, this study proposes an improved Faster R-CNN-based region of interest (ROI) detection algorithm, PPA-FRCNN, to provide high-quality regional input for fine pancreas segmentation. The algorithm replaces the traditional VGG-16 with ResNet50 as the backbone network, enhancing deep feature representation through residual connections. A dual-pyramid feature network structure is introduced to construct bidirectional feature aggregation paths from bottom-up and top-down directions, achieving cross-level multi-scale fusion of shallow edge textures and deep semantic information. A multi-branch multi-attention joint enhancement mechanism is designed to collaboratively strengthen pancreas feature representation from channel, context, and spatial dimensions. RoI Pooling is replaced with RoI Align to eliminate pixel quantization errors and optimize pancreas bounding box regression accuracy. Experimental results on the public NIH-TCIA dataset and clinical private dataset demonstrate that PPA-FRCNN achieves mAP50 of 89.7% and 90.7% respectively, with precision and recall rates superior to traditional Faster R-CNN, YOLOv7, and other comparative models, enabling high-precision localization of pancreas ROI and effectively reducing background interference and computational redundancy for subsequent segmentation tasks.