[Paper Review] 3D Anchor-Free Lesion Detector on Computed Tomography Scans

Ning Zhang, Dechun Wang+5 more|arXiv (Cornell University)|Aug 29, 2019

Medical Imaging Techniques and Applications21 references4 citations

TL;DR

This paper proposes a 3D anchor-free lesion detection framework for CT scans, modeling lesions as single keypoints instead of anchor-based bounding boxes. By eliminating anchors, the method achieves superior accuracy and faster inference speed compared to anchor-based baselines, particularly on challenging, large, or boundary-located lesions.

ABSTRACT

Lesions are injuries and abnormal tissues in the human body. Detecting lesions in 3D Computed Tomography (CT) scans can be time-consuming even for very experienced physicians and radiologists. In recent years, CNN based lesion detectors have demonstrated huge potentials. Most of current state-of-the-art lesion detectors employ anchors to enumerate all possible bounding boxes with respect to the dataset in process. This anchor mechanism greatly improves the detection performance while also constraining the generalization ability of detectors. In this paper, we propose an anchor-free lesion detector. The anchor mechanism is removed and lesions are formalized as single keypoints. By doing so, we witness a considerable performance gain in terms of both accuracy and inference speed compared with the anchor-based baseline

Motivation & Objective

• To address the limitations of anchor-based 3D lesion detection in CT scans, including high false positive rates and sensitivity to anchor configuration.
• To improve detection performance and inference speed by removing the anchor mechanism, which constrains generalization and increases computational load.
• To enable robust lesion detection in key slice-agnostic scenarios, enhancing practicality for real-world clinical applications.
• To investigate whether lesion detection can be effectively reformulated as a keypoint detection task in 3D, leveraging the non-overlapping nature of lesions in CT scans.

Proposed method

• The network uses a U-Net-like 3D architecture with skip connections between encoder and decoder pathways to preserve spatial details.
• Lesions are modeled as single keypoints (center points), eliminating the need for anchor templates and reducing false positives.
• Detection heads are attached to feature maps at multiple scales, predicting keypoint locations and confidence scores without bounding box regression.
• The model employs feature fusion via concatenation of upstream and downstream features at each resolution level to enhance localization accuracy.
• Training uses random crops of size 64×128×128 during training and sliding window inference during testing to handle large CT volumes.
• A modified loss function is used to supervise keypoint localization, with IoU-based matching between predicted and ground truth centers.

Experimental results

Research questions

• RQ1 Can a 3D anchor-free object detection framework achieve better performance than anchor-based methods in lesion detection on CT scans?
• RQ2 How does the anchor-free design affect inference speed and robustness to large or boundary-located lesions?
• RQ3 Does modeling lesions as single keypoints outperform anchor-based regression in terms of recall and precision on diverse lesion types?
• RQ4 How does the method perform in key slice-agnostic detection settings, compared to prior methods requiring prior knowledge of the key slice?

Key findings

• The anchor-free 3D lesion detector outperforms the anchor-based baseline in both FROC score and inference speed, with a 12% faster inference time on a Tesla K80.
• The anchor-free model shows significantly better robustness to large lesions (>30mm), with no performance degradation, while the anchor-based model suffers from instability and slow convergence.
• On the DeepLesion dataset, the anchor-free model achieves a mean FROC score of 0.852, surpassing the anchor-based model’s 0.821.
• The method maintains strong performance across lesion types, with minimal drop in recall for soft tissue lesions, unlike prior methods that degrade significantly on bone and kidney lesions.
• The model generalizes well to unseen lesion sizes and locations, including those near scan boundaries, due to the absence of anchor-induced localization bias.
• The absence of Non-Maximum Suppression (NMS) bottlenecks in the anchor-free design contributes to faster inference, especially in high-density regions.