[论文解读] Cancer Metastasis Detection With Neural Conditional Random Field

Yi Li, Wei Ping|arXiv (Cornell University)|Jun 19, 2018

AI in cancer detection参考文献 20被引用 79

一句话总结

NCRF 在 CNN patch 特征之上集成了一个全连接的 CRF，用于建模整个滑片图像（WSIs）补丁的空间相关性，从而提高转移灶检测并在 Camelyon16 上实现最先进的 FROC。

ABSTRACT

Breast cancer diagnosis often requires accurate detection of metastasis in lymph nodes through Whole-slide Images (WSIs). Recent advances in deep convolutional neural networks (CNNs) have shown significant successes in medical image analysis and particularly in computational histopathology. Because of the outrageous large size of WSIs, most of the methods divide one slide into lots of small image patches and perform classification on each patch independently. However, neighboring patches often share spatial correlations, and ignoring these spatial correlations may result in inconsistent predictions. In this paper, we propose a neural conditional random field (NCRF) deep learning framework to detect cancer metastasis in WSIs. NCRF considers the spatial correlations between neighboring patches through a fully connected CRF which is directly incorporated on top of a CNN feature extractor. The whole deep network can be trained end-to-end with standard back-propagation algorithm with minor computational overhead from the CRF component. The CNN feature extractor can also benefit from considering spatial correlations via the CRF component. Compared to the baseline method without considering spatial correlations, we show that the proposed NCRF framework obtains probability maps of patch predictions with better visual quality. We also demonstrate that our method outperforms the baseline in cancer metastasis detection on the Camelyon16 dataset and achieves an average FROC score of 0.8096 on the test set. NCRF is open sourced at https://github.com/baidu-research/NCRF.

研究动机与目标

• Motivate accurate detection of lymph node metastasis in whole-slide images (WSIs) for breast cancer diagnosis.
• Address block-wise patch classification limitations by modeling spatial correlations among neighboring patches.
• Propose an end-to-end trainable NCRF framework that jointly learns CNN features and CRF potentials.
• Demonstrate improved metastasis detection performance and smoother probability maps compared to patch-only baselines.

提出的方法

• Extract patch embeddings with ResNet-18 or ResNet-34 as the CNN feature extractor.
• Model conditional distribution over patch labels with a fully connected CRF on top of patch embeddings.
• Define unary potentials as negative log-likelihoods (CNN logits) and pairwise potentials as trainable weighted cosine distance between embeddings.
• Perform mean-field approximate inference to obtain patch-wise marginal distributions for end-to-end backpropagation.
• Train end-to-end using standard backpropagation with a small GPU overhead for the CRF mean-field iterations.
• Evaluate using probability maps, patch-wise accuracy, and average FROC on Camelyon16.

实验结果

研究问题

• RQ1 Can a neural CRF on patch embeddings improve the spatial coherence and accuracy of metastasis probability maps in WSIs?
• RQ2 Does end-to-end training of CNN features with a CRF lead to better patch-level classification than a CNN alone?
• RQ3 What is the impact of NCRF on metastasis detection performance (FROC) on the Camelyon16 dataset?

主要发现

Architecture	Baseline	NCRF	Architecture	Baseline	NCRF
ResNet-18	0.9242 ± 0.0007	0.9296 ± 0.0013
ResNet-34			0.9251 ± 0.0007	0.9338 ± 0.0014

• NCRF produces smoother probability maps with sharper boundaries than a patch-independent baseline.
• Patch classification accuracy improves when training the CNN with the CRF component (ResNet-18: 0.9296 vs 0.9242; ResNet-34: 0.9338 vs 0.9251).
• On Camelyon16 test set, NCRF improves average FROC score over the baseline (ResNet-18: 0.7934 vs 0.7825; ResNet-34: 0.7704 vs 0.7444).
• Best reported NCRF result: average FROC 0.8096 with ResNet-18.