Label-Free Liver Tumor Segmentation-CSDN博客

Qixin Hu1

Yixiong Chen2

Junfei Xiao3

Shuwen Sun4

Jieneng Chen3

Alan Yuille3

Zongwei Zhou3,*

1Huazhong University of Science and Technology

2The Chinese University of Hong Kong – Shenzhen

3 Johns Hopkins University

4The First Affiliated Hospital of Nanjing Medical University

Code and Visual Turing Test: https://github.com/MrGiovanni/SyntheticTumors

Abstract

Our synthetic tumors have two intriguing advantages: (I) realistic in shape and texture, which even medical professionals can confuse with real tumors; (II) effective for training AI models, which can perform liver tumor segmentation similarly to the model trained on real tumors—this result is exciting because no existing work, using synthetic tumors only, has thus far reached a similar or even close performance to real tumors.

我们的合成肿瘤有两个有趣的优势： (I)在形状和纹理上逼真，即使是医学专业人员也会与真实的肿瘤混淆；（II）训练AI模型有效，它可以执行肝脏肿瘤分割的真实肿瘤——这个结果令人兴奋，因为目前没有现有的工作，只使用合成肿瘤，迄今为止已经达到与真实肿瘤相似甚至接近的性能。

Moreover, our synthetic tumors can automatically generate many examples of small

(or even tiny) synthetic tumors and have the potential to improve the success rate of detecting small liver tumors, which is critical for detecting the early stages of cancer. In addition to enriching the training data, our synthesizing strategy also enables us to rigorously assess the AI robustness

此外，我们的合成肿瘤可以自动产生许多小的（甚至是微小的）合成肿瘤的例子，并有潜力提高检测小肝肿瘤的成功率，这对于检测癌症的早期阶段至关重要。除了丰富训练数据外，我们的综合策略还使我们能够严格评估人工智能的鲁棒性

Introduction

简介

In this paper, we handcraft a strategy to synthesize liver tumors in abdominal CT scans. Our

key novelties include (i) location without collision with vessels, (ii) texture with scaled-up Gaussian noise, and (iii) shape generated from distorted ellipsoids. These three aspects are proposed according to the clinical knowledge of liver tumors (detailed in §3.2). The resulting synthetic tumors are realistic—even medical professionals usually confuse them with real tumors in the visual examination (Figure 1; Table 2). In addition, the model trained on our synthetic tumors achieves a Dice Similarity Coefficient (DSC) of 59.81% for segmenting real liver tumors, whereas AI trained on real tumors obtains a DSC of 57.63% (Figure 2), showing that synthetic tumors have the potential to be used as an alternative to real tumors in training AI models.

在这篇论文中，我们手工制作了一种在腹部CT扫描中合成肝脏肿瘤的策略。我们的关键新颖性包括(i)不与血管碰撞的位置，（ii）具有放大高斯噪声的纹理，以及（iii）由扭曲的椭球体产生的形状。这三个方面是根据肝肿瘤的临床知识提出的（详见3.2）。由此产生的合成肿瘤是真实的——即使是医学专业人员在视觉检查中也会将其与真实的肿瘤混淆（图1；表2）。此外，模型训练我们的合成肿瘤实现骰子相似系数（DSC）的59.81%的分割真正的肝脏肿瘤，而人工智能训练真正的肿瘤获得DSC的57.63%（图2），表明合成肿瘤有可能被用作替代真正的肿瘤在训练人工智能模型。

The key contribution of ours is a synthetic tumor generator, which offers five advantages as summarized below.

1. The synthesis strategy embeds medical knowledge into an executable program, enabling the generation of realistic tumors through the collaboration of radiologists and computer scientists (§5.1; Table 2; Figure 3).

2. The entire training stage requires no annotation cost, and the resulting model significantly outperforms previous unsupervised anomaly segmentation approaches and tumor synthesis strategies (§5.2; Table 3).

3. The AI model trained on synthetic tumors can achieve similar performance to AI models trained on real tumors with per-voxel annotation in real tumors segmentation, and can be generalized to CT scans with healthy liver and scans from other hospitals (§5.3; Figure 4).

4. The synthesis strategy can generate a variety of tumors for model training, including those at small, medium, and large scales, and therefore have the potential to detect small tumors and facilitate the early detection of liver cancer (§5.4; Figure 5).

5. The synthesis strategy allows for straightforward manipulation of parameters such as tumor location, size, texture, shape, and intensity, providing a comprehensive test-bed for evaluating AI models under out-of distribution scenarios (§5.5; Figure 6).

1.该综合策略将医学知识嵌入到一个可执行的程序中，通过放射科医生和计算机科学家的合作，能够生成真实的肿瘤（5.1；表2；图3）。

2.整个训练阶段不需要注释成本，所得到的模型明显优于以前的无监督异常分割方法和肿瘤合成策略（5.2；表3）。

3.在合成肿瘤上训练的AI模型在真实肿瘤分割中可以达到与在真实肿瘤上训练的每体素注释的人工智能模型相似的性能，并且可以推广到健康肝脏的CT扫描和其他医院的扫描（5.3；图4）。

4.该综合策略可以产生多种用于模型训练的肿瘤，包括小、中、大规模的肿瘤，因此具有检测小肿瘤和促进肝癌早期发现的潜力（5.4；图5）。

5.该合成策略允许直接操纵肿瘤的位置、大小、纹理、形状和强度等参数，为在非分布场景下评估AI模型提供了一个全面的试验台（5.5；图6）。

Method

方式

Tumor Generation

To localize the liver, we first apply the pre-trained nnUNet1 to the CT scans. With a coarse location of the liver available, we then develop a sequence of morphological image-processing operations to synthesize realistic tumors within the liver (see Figure 3). The tumor generation consists of four steps: (1) location selection, (2) texture generation, (3) shape generation, and (4) post-processing

为了定位肝脏，我们首先将预先训练好的nnUNet1应用于CT扫描。利用肝脏的粗糙位置，我们开发了一系列形态学图像处理操作，以合成肝脏内的真实肿瘤（见图3）。肿瘤的形成包括四个步骤：(1)位置选择、(2)纹理生成、(3)形状生成和(4)后处理

1、位置选择：

The first step is to select a proper location for the tumor. This step is crucial because liver tumors usually do not allow any vessels (e.g., hepatic vein, portal vein, and inferior vena cava) to pass through them. To avoid the blood vessels, we first conduct vessel segmentation through the voxel value thresholding.

第一步是为肿瘤选择一个合适的位置。这一步是至关重要的，因为肝肿瘤通常不允许任何血管（如肝静脉、门静脉和下腔静脉）通过它们。为了避免血管的发生，我们首先通过体素值阈值法进行血管分割。

大致就是先将可能存在血管的区域分割开来，然后避免生成病灶侵入血管的区域。

2、纹理生成

The HU values of liver and tumor textures follow the Gaussian distributions. To obtain realistic

tumor textures, we first generate a 3D Gaussian noise with the predefined mean HU intensity µt and the same standard deviation σp as the hepatic parenchyma (the liver area excluding vessels), T(x, y, z) ∼ N (µt, σp). Since the random Gaussian noise is usually too sharp as the texture for the tumor, we soften the texture by scaling it up with spline interpolation of the order 3 (cubic interpolation) on x, y, z directions. The scaled-up texture is denoted as T ′ (x, y, z) in this work, we want it exhibits graininess close to the hepatic parenchyma. The scaling factor η ∈ [1, ∞) determines how rough the generated grain feels. η = 1 means the Gaussian texture is not scaled, resulting in large value fluctuation between adjacent voxels.

3、形状生成

4、后处理

分界线+++++++++++++++++++++++++++++++++++++++