International Journal of Computer Assisted Radiology and Surgery
Evolutionary image simpliﬁcation for lung nodule classiﬁcation with
convolutional neural networks
· Gabriele von Voigt
Received: 11 January 2018 / Accepted: 17 May 2018
© CARS 2018
Purpose Understanding decisions of deep learning techniques is important. Especially in the medical ﬁeld, the reasons for
a decision in a classiﬁcation task are as crucial as the pure classiﬁcation results. In this article, we propose a new approach to
compute relevant parts of a medical image. Knowing the relevant parts makes it easier to understand decisions.
Methods In our approach, a convolutional neural network is employed to learn structures of images of lung nodules. Then,
an evolutionary algorithm is applied to compute a simpliﬁed version of an unknown image based on the learned structures
by the convolutional neural network. In the simpliﬁed version, irrelevant parts are removed from the original image.
Results In the results, we show simpliﬁed images which allow the observer to focus on the relevant parts. In these images,
more than 50% of the pixels are simpliﬁed. The simpliﬁed pixels do not change the meaning of the images based on the
learned structures by the convolutional neural network. An experimental analysis shows the potential of the approach. Besides
the examples of simpliﬁed images, we analyze the run time development.
Conclusions Simpliﬁed images make it easier to focus on relevant parts and to ﬁnd reasons for a decision. The combination
of an evolutionary algorithm employing a learned convolutional neural network is well suited for the simpliﬁcation task. From
a research perspective, it is interesting which areas of the images are simpliﬁed and which parts are taken as relevant.
Keywords Understanding deep learning · Image simpliﬁcation · Evolutionary algorithm · Convolutional neural networks
The amount of medical images is dramatically growing .
This leads to various challenges, for example, in the ﬁeld of
data storage and sharing  to make data available when
they are required. When medical data are required, mostly,
not the data itself are interesting. The information within the
medical data is crucial. Thus, the determination of required
information from medical image data is a key concern. In the
past, the task of a medical doctor to determine the required
information from medical image data was primary . But
as the amount of medical images is growing consistently,
Electronic supplementary material The online version of this article
(https://doi.org/10.1007/s11548-018-1794-7) contains supplementary
material, which is available to authorized users.
Computational Health Informatics, Leibniz University
Hanover, Schloßwender Str. 5, 30159 Hanover, Germany
automatic and semiautomatic algorithms are needed to under-
take this task.
In the ﬁeld of image classiﬁcation, there was tremen-
dous progress in the last past years . New deep learning
techniques make it possible to achieve highly accurate clas-
siﬁcation scores [15,40]; i.e., they can predict an information
based on an image. In the medical ﬁeld, this could be a com-
puter tomography image of lung nodules and then, a deep
learning technique predicts whether the lung nodules are
malignant or benign . The state-of-the-art approach from
the ﬁeld of deep learning for image classiﬁcation is convo-
lutional neural networks (CNNs) . We introduce CNNs
more in detail in Sect. 2.3. They are a very advanced classi-
ﬁcation method but unfortunately due to their complexity, it
is very challenging to determine the exact reasons for their
classiﬁcation decisions .
Reasoning decisions is verycrucial in the medical ﬁeld .
For a medical decision support system, the reasoning might
be more important than the actual decision. Medical doctors
take decisions by themselves as they are responsible for their
decisions. Thus, they are more interested in arguments and