<![CDATA[Pleural effusion detection from chest X-ray images remains challenging due to variations in lesion scale and severe foreground-background imbalance within two-stage object detectors. This study empirically investigates the effects of anchor representation and Region Proposal Network (RPN) sampling on the detection process of Faster R-CNN for pleural effusion detection. A Faster R-CNN with ResNet-50 and Feature Pyramid Network (FPN) was evaluated using representative anchors derived from ground-truth object characteristics and adjusted RPN Intersection over Union (IoU) assignment thresholds. The analysis examined anchor representation, RPN sample distribution, proposal quality, detection confidence, and stage-wise detection outcomes. Representative anchors improved geometric correspondence with the target objects, while RPN threshold adjustment substantially increased positive RPN samples from approximately 4 to 110 per image. Proposal analysis showed high spatial accessibility, with proposal IoU coverage at IoU ≥ 0.50 reaching up to 99.35% across the evaluated configurations. However, stage-wise analysis showed that classification was the dominant source of detection failure, accounting for 15.03–16.99% of the evaluated objects, while proposal and localization failures remained comparatively limited. These findings demonstrate that optimization in anchor representation and RPN sampling can enhance upstream proposal characteristics, but such improvements do not necessarily translate proportionally into final detection performance.]]>
Copyrights © 2026
