PDSE: A Multiple Lesion Detector for CT Images Using PANet and Deformable Squeeze-and-Excitation Block

2025-06-04

Annual Conference on Medical Image Understanding and Analysis

AI Summary

Detecting lesions in CT images remains challenging due to substantial variability in lesion type, scale, and spatial location. To address this, we propose PDSE, a one-stage detection framework built upon RetinaNet. PDSE incorporates PANet to enhance low-level feature representation and introduces a novel deformable Squeeze-and-Excitation (dSE) module that enables geometrically adaptive channel-wise attention modeling. Furthermore, it fuses multi-scale features with channel-adaptive weighting to improve contextual discrimination. This design significantly boosts detection sensitivity for small and multi-scale lesions. Evaluated on the DeepLesion benchmark, PDSE achieves an mAP exceeding 0.20—setting a new state-of-the-art among comparable one-stage methods at the time. The results empirically validate the effectiveness of jointly enhancing low-level features and incorporating geometry-aware channel attention for lesion detection in CT imaging.

Technology Category

• Computer Vision: Multi-modal Vision
• Search and Optimization: Learning to Search
• Machine Learning: Multimodal Learning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Security and Privacy: Privacy-enhancing technologies
• Responsible Web: Data and user privacy-enhancing technologies for the Web

Abstract

Detecting lesions in Computed Tomography (CT) scans is a challenging task in medical image processing due to the diverse types, sizes, and locations of lesions. Recently, various one-stage and two-stage framework networks have been developed to focus on lesion localization. We introduce a one-stage lesion detection framework, PDSE, by redesigning Retinanet to achieve higher accuracy and efficiency for detecting lesions in multimodal CT images. Specifically, we enhance the path aggregation flow by incorporating a low-level feature map. Additionally, to improve model representation, we utilize the adaptive Squeeze-and-Excitation (SE) block and integrate channel feature map attention. This approach has resulted in achieving new state-of-the-art performance. Our method significantly improves the detection of small and multiscaled objects. When evaluated against other advanced algorithms on the public DeepLesion benchmark, our algorithm achieved an mAP of over 0.20.

Problem

• Detecting diverse lesions in CT scans accurately
• Improving lesion localization with one-stage framework PDSE
• Enhancing small and multiscaled object detection performance

Innovation

• Redesigned Retinanet for lesion detection
• Enhanced path aggregation with low-level features
• Adaptive SE block with channel attention