[论文解读] Automatic Liver Lesion Detection using Cascaded Deep Residual Networks

Lei Bi, Jin‐Man Kim+2 more|arXiv (Cornell University)|Apr 10, 2017

Brain Tumor Detection and Classification参考文献 14被引用 84

一句话总结

本文提出级联深度残差网络（ResNet）与多尺度融合，用于自动从CT分割肝脏及肝病变，达到先进水平并在LiTS挑战中获得第四名。

ABSTRACT

Automatic segmentation of liver lesions is a fundamental requirement towards the creation of computer aided diagnosis (CAD) and decision support systems (CDS). Traditional segmentation approaches depend heavily upon hand-crafted features and a priori knowledge of the user. As such, these methods are difficult to adopt within a clinical environment. Recently, deep learning methods based on fully convolutional networks (FCNs) have been successful in many segmentation problems primarily because they leverage a large labelled dataset to hierarchically learn the features that best correspond to the shallow visual appearance as well as the deep semantics of the areas to be segmented. However, FCNs based on a 16 layer VGGNet architecture have limited capacity to add additional layers. Therefore, it is challenging to learn more discriminative features among different classes for FCNs. In this study, we overcome these limitations using deep residual networks (ResNet) to segment liver lesions. ResNet contain skip connections between convolutional layers, which solved the problem of the training degradation of training accuracy in very deep networks and thereby enables the use of additional layers for learning more discriminative features. In addition, we achieve more precise boundary definitions through a novel cascaded ResNet architecture with multi-scale fusion to gradually learn and infer the boundaries of both the liver and the liver lesions. Our proposed method achieved 4th place in the ISBI 2017 Liver Tumor Segmentation Challenge by the submission deadline.

研究动机与目标

• 在 CT 图像中推动自动、准确的肝脏病变分割，用于 CAD/CDS。
• 通过使用深度残差网络进行区分性特征学习来克服浅层 FCN 的局限性。
• 提出一个级联 ResNet 架构以迭代地细化肝脏和病变边界。
• 结合多尺度融合以应对各中心分辨率差异。

提出的方法

• 通过将深度残差网络转换为具有上采样和空洞卷积的类 FCN 模型来进行分割。
• 构建一个级联 ResNet 框架，在训练和测试中利用前一轮的概率图来细化肝脏和病变分割。
• 应用多尺度输入重新采样并对输出取平均以获得最终预测（多尺度融合）。
• 通过 Hounsfield 单位窗宽化和归一化到 [0,1] 对 CT 进行预处理。
• 以 ImageNet 的微调开始训练，随后进行领域特定微调，结合数据增强和 SGD 优化。

实验结果

研究问题

• RQ1 基于 ResNet 的架构能否超越基于 VGGNet 的 FCN 在 CT 图像中对肝脏和肝病变分割的表现？
• RQ2 带有迭代细化的级联 ResNet 是否能提升肝脏和病变分割边界的准确性？
• RQ3 多尺度融合是否在来自不同中心的 CT 研究中提供稳健的性能？
• RQ4 在级联 ResNet 分割中加入 3D-CRF 后处理的影响是什么？

主要发现

• 级联 ResNet 在肝脏与病变分割方面优于基于 VGG 的 FCN。
• 未进行后处理的级联 ResNet 获得 Dice 95.51%（肝脏）和 49.83%（病变）；Jaccard 91.45%（肝脏）和 38.59%（病变）。
• 结合多尺度融合的级联 ResNet 获得最佳结果：Dice 95.90%（肝脏）和 50.01%（病变）；Jaccard 92.19%（肝脏）和 38.79%（病变）。
• 3D-CRF 后处理相比基础级联 ResNet 降低了病变分割性能。
• 多尺度融合对不同研究分辨率和中心具有鲁棒性，有助于 LiTS 提交中的高排名（第4 名）。