[paper reading] Ca net: leveraging contextual features for lung cancer prediction

[ Address of thesis ] [ Code ] [MICCAI 21]

Abstract

In the early diagnosis of lung cancer , An important step is to malignant each pulmonary nodule / Benign classification . For this classification , Characteristics of nodules ( Like shape 、 edge ) Has always been the main focus . lately , Background features have attracted more and more attention because of the supplementary information they provide . In the clinical , This background feature refers to the characteristics of the structure around the nodule , such ( Together with the characteristics of nodules ) They can expose malignant / Benign discrimination patterns , Such as blood vessel convergence and fissure attachment . In order to take advantage of these contextual features , We propose a context aware network (CA-Net), It can extract nodule and context features , Then in malignant / Effectively integrate them in benign classification . In order to accurately identify the twisted nodules / Contextual features of attached structures , We use the characteristics of nodules as a reference through the attention mechanism . Besides , We propose a feature fusion module , The weights of the features of each nodule and the background features can be adjusted adaptively . The practicability of our method is better than Kaggle Data science bowl in the competition 2017 The first place in the annual data set has obvious advantages .

Method

This article is a relatively advanced context attention Mechanism , That is, the distortion of pulmonary nodules is introduced , Information like attachment . The specific process is as follows ：

Detection of lung disease is essentially a two-stage classification problem . First , Use a nodule test (Nodule Detection) The algorithm puts CT The lung nodules summarized in the image are detected , This article directly uses a TNNLS Classic algorithm on ; next , To classify each nodule (Nodule Malignancy Classification), The probability of judging it as malignant ; Last , According to the state of each nodule, the lung CT State for comprehensive diagnosis (Cancer Prediction). Diagnosis is actually a summary of prediction probabilities ：$$P(Y\mid I)=1-\left(1-p_l\right)\prod _{k=1}^K\left(1-P\left(M\mid I_{N_k}\right)\right)$$ Therefore, the core work of this paper is actually on the classification of pulmonary nodules , There are three steps ： feature extraction , Contextual attention and feature fusion . Feature extraction is actually using 3D-UNet Get feature map.

Context attention ：

What this step does is how to transform the original features $X$ Extract nodule features respectively $X_N$ And contextual features $X_C$.$X_N$ It's easy to get , Just do one ROI Pooling You can extract . The highlight here is $X_C$ Acquisition needs $X_N$ To assist .

As you can see from the diagram ,$X_C$ Is in $X_S$ On the basis of doing a spatial attention get ,$X_S=X-X_C$. As for this spatial attention , In fact, it is caused by $X_N$ Made a "Nodule Encoding",$X_S$ Made a "Surrounding Encoding", Put the results obtained concat Get up and do attention map. As for these two Encoding block, This article does not point out its structure , However, follow the general practice of attention to use any learnable network components （ Even full connection layer ） It should all be possible .

Feature fusion ：

As for the fusion of this feature, it is also the thought of spatial attention , It's just that it's even simpler and rougher here , That is to get ：$$X_{\text{fuse }}=\mathrm{Concat}\left({\omega }_N\times X_N,{\omega }_C\times X_C\right)$$ So I directly put on a small MLP( It can be considered as several full connection layers ) To calculate this coefficient $\omega$.