[论文解读] Towards Single-phase Single-stage Detection of Pulmonary Nodules in Chest CT Imaging

Zhongliu Xie|arXiv (Cornell University)|Jul 16, 2018

Lung Cancer Diagnosis and Treatment被引用 8

一句话总结

该论文提出了一种单阶段、单阶段的深度学习框架，用于在3D胸部CT扫描中端到端检测肺结节，采用改进的ResNet-FPN架构，结合RReLU激活函数和一种新颖的基于锚点的补丁采样策略，以解决极端类别不平衡问题。该方法在[1, 2, 4]次扫描中每幅图像的假阳性数下实现了SOTA性能，FROC得分为0.9767，显著优于以往的两阶段方法，在准确率和推理速度方面均表现出色（每幅扫描约35秒）。

ABSTRACT

Detection of pulmonary nodules in chest CT imaging plays a crucial role in early diagnosis of lung cancer. Manual examination is highly time-consuming and error prone, calling for computer-aided detection, both to improve efficiency and reduce misdiagnosis. Over the years, a range of systems have been proposed, mostly following a two-phase paradigm with: 1) candidate detection, 2) false positive reduction. Recently, deep learning has become a dominant force in algorithm development. As for candidate detection, prior art was mainly based on the two-stage Faster R-CNN framework, which starts with an initial sub-net to generate a set of class-agnostic region proposals, followed by a second sub-net to perform classification and bounding-box regression. In contrast, we abandon the conventional two-phase paradigm and two-stage framework altogether and propose to train a single network for end-to-end nodule detection instead, without transfer learning or further post-processing. Our feature learning model is a modification of the ResNet and feature pyramid network combined, powered by RReLU activation. The major challenge is the condition of extreme inter-class and intra-class sample imbalance, where the positives are overwhelmed by a large negative pool, which is mostly composed of easy and a handful of hard negatives. Direct training on all samples can seriously undermine training efficacy. We propose a patch-based sampling strategy over a set of regularly updating anchors, which narrows sampling scope to all positives and only hard negatives, effectively addressing this issue. As a result, our approach substantially outperforms prior art in terms of both accuracy and speed. Finally, the prevailing FROC evaluation over [1/8, 1/4, 1/2, 1, 2, 4, 8] false positives per scan, is far from ideal in real clinical environments. We suggest FROC over [1, 2, 4] false positives as a better metric.

研究动机与目标

• 消除肺结节检测中传统的两阶段范式，该范式将候选结节检测与假阳性减少分离。
• 解决3D结节检测中极端的类间和类内类别不平衡问题，其中正样本远多于易分的负样本。
• 开发一种端到端的、单网络解决方案，无需迁移学习或后处理即可同时执行分类和边界框回归。
• 提出一种临床上相关的评估指标——[1, 2, 4]次假阳性每幅扫描的FROC，更符合放射科医生的关注重点。
• 通过基于动态补丁的采样策略和自适应锚点，提升推理速度和训练效率。

提出的方法

• 采用改进的ResNet-FPN架构，通过跳跃连接和侧向连接增强3D体数据中的特征学习和多尺度检测能力。
• 采用RReLU激活函数以缓解Dying ReLU问题，并提升深层网络中的训练稳定性。
• 提出一种基于锚点的补丁采样策略，动态选择正样本和困难负样本，从而减轻来自巨大负样本池的训练负担。
• 通过定期更新锚点，确保采样在训练各周期中保持相关性，提升模型泛化能力。
• 在3D CT体数据上进行端到端训练，无需预训练或迁移学习，直接回归结节位置和概率。
• 采用结合分类头和回归头的损失函数，以同时优化检测准确率和定位精度。

实验结果

研究问题

• RQ1 单阶段、单阶段的深度学习模型是否能在3D CT扫描的肺结节检测中超越现有两阶段流水线？
• RQ2 基于锚点的补丁采样策略在3D结节检测的端到端训练中，对缓解极端类别不平衡问题的有效性如何？
• RQ3 与最先进的两阶段方法相比，所提方法是否实现了更优的检测性能和更快的推理速度？
• RQ4 [1, 2, 4]次假阳性每幅扫描的FROC是否比传统的[1/8, 1/4, 1/2, 1, 2, 4, 8]尺度更具临床意义？
• RQ5 该模型是否能在不依赖迁移学习或后处理的前提下，实现高敏感性并保持极低的假阳性数？

主要发现

• 所提出的单模型在[1, 2, 4]次假阳性每幅扫描的新评估标准下，取得了0.9767的FROC得分，显著优于以往方法。
• 集成模型在同一指标下达到0.9817的FROC得分，表明其具备出色的泛化能力和鲁棒性。
• 在标准基准上的FROC[luna]得分为0.9351，超越了以往的SOTA方法。
• 推理时间约为每幅CT扫描35秒，远快于两阶段方法。
• 基于锚点的补丁采样策略通过仅聚焦于困难负样本和正样本，显著减少了训练样本数量，提升了训练效率。
• 放射科医生的临床反馈支持采用[1, 2, 4]次假阳性每幅扫描作为更实用的评估指标，因其更关注减少假阴性而非假阳性。