Conditional Training with Bounding Map for Universal Lesion Detection

Han Li, Long Chen, Hu Han, Ying Chi, S. Kevin Zhou

Abstract

Universal Lesion Detection (ULD) in computed tomography plays an essential role in computer-aided diagnosis. Promising ULD results have been reported by coarse-to-fine two-stage detection approaches, but such two-stage ULD methods still suffer from issues like imbalance of positive v.s. negative anchors during object proposal and insufficient supervision problem during localization regression and classification of the region of interest (RoI) proposals. While leveraging pseudo segmentation masks such as bounding map (BM) can reduce the above issues to some degree, it is still an open problem to effectively handle the diverse lesion shapes and sizes in ULD. In this paper we propose a BM-based conditional training for two-stage ULD, which can (i) reduce positive vs. negative anchor imbalance via a BM-based conditioning (BMC) mechanism for anchor sampling instead of traditional IoU-based rule; and (ii) adaptively compute size-adaptive BM (ABM) from lesion bounding-box, which is used for improving lesion localization accuracy via ABM-supervised segmentation. Experiments with four state-of-the-art methods show that the proposed approach can bring an almost free detection accuracy improvement without requiring expensive lesion mask annotations.

Introduction

Universal Lesion Detection (ULD) is a critical component of computer-aided diagnosis (CAD) systems for CT scans. However, ULD faces unique challenges compared to natural image detection. Lesions vary significantly in size, shape, and appearance, and their occurrence in CT slices is sparse. Most state-of-the-art systems employ a two-stage detection framework (e.g., Faster R-CNN), which first generates region proposals (anchors) and then refines them.

Two primary bottlenecks hinder these systems in medical imaging:

  1. Anchor Imbalance: Standard detection rules require high Intersection over Union (IoU) overlap to label an anchor as positive. In CT scans, where lesions are rare, this leads to an extreme imbalance (up to 200:1) between negative and positive samples, making it difficult for models to learn effectively.
  2. Insufficient Supervision: The second stage often relies on a simple binary classification score, which lacks the spatial detail needed to distinguish subtle lesions from similar-looking healthy tissue.
Comparison of anchor sampling strategies: (a) traditional IoU-based method, (b) standard bounding map, and (c) the proposed BMC method using soft label maps.

Proposed Methodology

The researchers propose a Conditional Training with Bounding Map framework to address these issues. The framework introduces three key innovations:

BM-based Conditioning (BMC)

Instead of a rigid IoU threshold, BMC utilizes a "per-pixel objectness map." An anchor is classified as positive if its center falls within a region of the Bounding Map (a soft-label map with peak values at the lesion center) that exceeds a predefined threshold. This approach significantly increases the number of high-quality positive samples available for training.

Size-Adaptive Bounding Map (ABM)

Standard bounding maps can be suboptimal for lesions of varying scales. Small lesions may provide insufficient signal, while large ones can dominate the loss function. The proposed ABM adaptively "enhances" the signal for small lesions and "weakens" it for larger ones, ensuring balanced supervision across all scales.

ABM-Supervised Segmentation

To provide richer supervision, the framework includes a segmentation branch that uses the ABM as a "pseudo-mask." This forces the network to learn the spatial distribution of lesions without requiring manual pixel-level annotations, which are costly to obtain.

The proposed two-stage framework: Stage-1 uses BMC for objectness and regression training, while Stage-2 incorporates ABM-supervised segmentation for refined localization.

Experimental Results

The method was evaluated on the DeepLesion dataset, which contains over 32,000 lesions with diameters ranging from 0.21mm to 342.5mm.

The proposed framework demonstrated significant performance gains across multiple state-of-the-art baselines:

  • Faster R-CNN: Average sensitivity improved from 70.85% to 77.67%.
  • AlignShift: Achieved a new state-of-the-art sensitivity of 86.83%.
  • Efficiency: The framework adds less than 10% to inference time, making it highly practical for clinical use.

Ablation studies highlighted that BMC objectness training was the most impactful component, providing a ~1.8% performance boost on its own.

Conclusion and Future Work

This work demonstrates that pseudo-labels derived from bounding boxes can significantly enhance ULD performance. By relaxing the IoU bottleneck and providing size-adaptive spatial supervision, the model achieves superior localization and detection accuracy. Future research may explore integrating 3D spatial context and refining the geometric approximations used in pseudo-masks to better reflect complex anatomical boundaries.

Summary updated

Abstract

Universal Lesion Detection (ULD) in computed tomography plays an essential role in computer-aided diagnosis. Promising ULD results have been reported by coarse-to-fine two-stage detection approaches, but such two-stage ULD methods still suffer from issues like imbalance of positive v.s. negative anchors during object proposal and insufficient supervision problem during localization regression and classification of the region of interest (RoI) proposals. While leveraging pseudo segmentation masks such as bounding map (BM) can reduce the above issues to some degree, it is still an open problem to effectively handle the diverse lesion shapes and sizes in ULD. In this paper we propose a BM-based conditional training for two-stage ULD, which can (i) reduce positive vs. negative anchor imbalance via a BM-based conditioning (BMC) mechanism for anchor sampling instead of traditional IoU-based rule; and (ii) adaptively compute size-adaptive BM (ABM) from lesion bounding-box, which is used for improving lesion localization accuracy via ABM-supervised segmentation. Experiments with four state-of-the-art methods show that the proposed approach can bring an almost free detection accuracy improvement without requiring expensive lesion mask annotations.
BibTeX
@inproceedings{li2021conditional,
  author = {Han Li and Long Chen and Hu Han and Ying Chi and S. Kevin Zhou},
  title = {{Conditional Training with Bounding Map for Universal Lesion Detection}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2021},
  url = {https://openreview.net/forum?id=rwMj7tba5O},
}
Figure