3D-CNN深度解析：医疗成像中3维卷积神经网络的全面综述
需积分: 9 3 浏览量
更新于2024-07-14
收藏 3.95MB PDF 举报
3D-CNN在医疗成像领域的应用：深度学习视角下的综述
随着机器学习技术的进步、图形处理能力的提升以及医疗影像数据的日益增多，3维卷积神经网络(3D Convolutional Neural Networks, 3D-CNN)在医学图像分析中的地位日益凸显。这篇综述论文由Satya P. Singh等人撰写，来自新加坡南洋理工大学的研究团队，涵盖了Lee Kong Chian医学院、认知神经影像中心等多个机构。他们的研究聚焦于3D-CNN在医学成像中的深度学习应用，尤其是在诊断支持、病理分析和疾病预测等方面。
3D-CNN相比于传统的2D CNN，其优势在于能够捕捉到图像的三维空间结构，这对于医疗领域至关重要，因为许多医学图像如MRI、CT和PET等都是三维数据。通过三维滤波器，3D-CNN可以更好地理解组织间的立体关系，这对于识别肿瘤、血管结构或解剖结构的精确分析具有显著优势。
论文详细讨论了3D-CNN在诸如病灶检测、分割、功能成像分析以及疾病风险评估等任务中的应用。这些应用包括但不限于神经系统疾病（如脑部肿瘤和阿尔茨海默病）、心脏病（如冠状动脉钙化检测）和肺部疾病（如CT扫描中的结节分析）。3D-CNN通过卷积层、池化层和全连接层的组合，实现了对复杂医学图像特征的高效提取和抽象表示，从而提高了诊断准确性和自动化水平。
此外，作者还探讨了3D-CNN模型的设计优化，如多尺度卷积、残差连接和注意力机制等技术，以应对医学图像数据的高维度和异质性挑战。同时，他们也提到了在实际应用中可能遇到的数据不足、标注难题和模型解释性问题，并对这些问题进行了相应的解决策略和未来研究方向的展望。
该综述论文不仅提供了对3D-CNN在医疗成像领域深度学习应用的全面概述，还为研究人员和临床医生提供了宝贵的学习资源，帮助他们了解如何有效地利用这种技术来改善疾病的早期诊断、个性化治疗和临床决策支持。随着技术的不断进步，3D-CNN在未来的医疗影像分析中有望发挥更大的作用。
Sensors 2020, 20, 5097 5 of 24
introduce sampling noise, which appears in training data sets but not in real test datasets even if
both are drawn from the same distribution. This scenario leads to overfitting and there have been
several strategies [
26
] to tackle the problem, such as early stopping of the training epochs and weight
penalties (L1 and L2 regularizations, soft weight sharing, and pooling). Ensemble models of several
CNNs with different configurations on the same dataset are known for their overfitting. However,
this leads to extra computational and maintenance cost for training several models. Moreover, training
a large network requires large datasets, but the availability of such datasets in the field of medical
imaging is very rare. Even if one can train large networks with a versatile setting of parameters,
testing these networks is not feasible in a real-time situation due to the nature of medical imaging
systems. In the case of ensemble models, a CNN model can also simulate multiple configurations just
by probabilistically dropping out edges and nodes. Dropout is a kind of regularization technique to
reduce overfitting by temporarily dropping a unit out of the network [
27
]. This simple idea shows a
significant improvement in CNN performance.
Batch normalization: The input of each hidden layer dynamically changes during training because
the parameters in the previous layer update at each training epoch. If these changes are large, the search
for an optimal hyperparameter becomes difficult for the network and may be computationally expensive
to reach an optimal value. This problem can be solved by an algorithm called batch normalization,
which was proposed by two researchers [
28
]. Batch normalization allows the use of a higher learning
rate and thereby achieves the optimal value in less time. It facilitates the smooth training of deeper
network architectures in less time. The normalization of data from a particular batch is about finding
the mean and variance of the data points from mini-batch and normalizing them to have a zero mean
and unit variance.
In backward pass, the CNN adjusts its weights and parameters according to the output by
calculating the error through some loss functions,
e
(other names are cost function and error function)
and backpropagating the error with some rules towards the input. The loss is calculated by taking the
partial derivative of
e
w.r.t., which is the output of each neuron in that layer, such as
∂e/y
`
i,j,k
for the
output,
y
`
i,j,k
of
(
i, j, k
)
th
unit in layer
`
. The cFhain rule allows us to write and add up the contribution
of each variable as follows:
∂e
∂x
`
i,j,k
=
∂e
∂y
`
i,j,k
∂ f
y
`
i,j,k
∂x
`
i,j,k
=
∂e
∂y
`
i,j,k
f
0
x
`
i,j,k
#. (3)
Weights in the previous convolutional layer can be updated by backpropagating the error to the
previous layer according to the following equation:
∂e
∂y
`−1
i,j,k
=
n1−1
X
a=0
n2−1
X
b=0
n3−1
X
c=0
∂e
∂x
`
(i−a),(j−b),(k−c)
∂x
`
(i−a),(j−b),(j−b)
∂y
`−1
i,j,k
. (4)
=
n1−1
X
a=0
n2−1
X
a=0
n3−1
X
b=0
∂e
∂x
`
(i−a),(j−b),(k−c)
ω
a,b,c
. (5)
Equation (5) allows us to calculate the error for the previous layer. Further, the above eq. makes
sense for those points which are n times away from each side of the input data. This situation can be
avoided by simply padding with zeros to the end of each side of the input volume.
2.2. Breakthroughs in CNN Architectural Advances
Several different versions of CNN have been proposed in the literature to improve model
performance. In 2011, Krizhevsky et al. [
14
] presented a deep CNN architecture. A systematic
architecture of AlexNet is shown in Figure 4. AlexNet has five convolutional layers and three fully
剩余23页未读，继续阅读
点击了解资源详情
142 浏览量
点击了解资源详情
2021-10-10 上传
111 浏览量
159 浏览量
2024-06-17 上传
2024-07-21 上传
123 浏览量
