Page 68 - AIH-2-2

Artificial Intelligence in Health Improved liver tumor segmentation with dense networks

tumor segmentation in CT images, named intra- and predefined CT window so as to increase the visualization of
inter-block densely connected FCN (I -DenseFCN). This certain pathological features relevant to their task. For liver
2
framework incorporates an adaptive learning threshold tumor segmentation, most previous works convert raw CT
processing method to enhance the distinguishability of images into grayscale images using a preset window setting
tumors from surrounding tissues, thereby improving the or directly utilize the full range of CT image intensity
16
generalization of segmentation capabilities on potential values without windowing input. Christ et al. carried out
samples compared to previous methods. To reduce the pre-processing with a CT window ranging [−100, 400]
computational cost associated with the multiscale tumor HU to exclude irrelevant organs and objects, followed by a
29
segmentation feature, we establish connections between histogram equalization operation. Li et al. truncated the
feature maps of different levels and propose a decoder image intensity values of all CT volumes to the range of
network with inter-block dense connections. Furthermore, [−200, 250] HU to remove the irrelevant details. Han set
30
we designed a dynamic weighting mechanism to address the HU value range to [−200, 200], whereas Bi et al. used
19
loss functions at different stages of model training. In [−160, 240] HU range. These CT image pre-processing
summary, the threefold contribution of this work is as settings are subjective and empirical and they vary from
follows: researcher to researcher. It is an open question whether
(i) We first propose a CT image pre-processing module that the preset window range is optimal for specific datasets,
is integrated and jointly trained with the segmentation models, and for the eventual segmentation performance.
network, allowing optimal pre-processing settings Despite the importance of optimal window settings in
through backpropagation, rather than separate pre- clinical practice, their impact on the model performance
processing with an empirically preset CT window. has been largely overlooked in the liver tumor segmentation
Our experimental result demonstrates the positive task. Inspired by the work of Lee et al., we propose a
31
efficacy of this module on segmentation performance CT image pre-processing module that is jointly trained
(ii) A decoder module with inter-block dense connections with the segmentation network to find optimal window
is built to effectively cope with the challenges posed settings by learning instead of predefining the value.
by multiscale tumors, resulting in further improved A fully end-to-end segmentation framework is generated
performance simultaneously. The pre-processing module is composed
(iii) We attempt to refine the segmentation of tumor of a 1 × 1 convolutional layer and an activation layer with
boundaries in a coarse-to-fine manner. Specifically, the sigmoid function. The module structure is shown in
a hybrid loss function combining the cross-entropy Figure 2.
and the Lovász-Softmax loss with dynamic weights
27
is developed to optimize our model. Experimental The proposed module could be formulated as follows:
results indicate that this strategy enhances the U
precision of tumor boundary segmentation. Fx Wx b (I)
1 e
2. Methods x is the raw intensity value in CT images. The constant U

In this section, we first introduce the CT image pre- is the upper bound of the pre-processing function. W and
processing module and then describe the segmentation b are the weight and bias of the 1 × 1 convolutional layer,
network in detail. The hybrid loss function design is respectively, and they are calculated using the equations of
presented in the end. The pipeline of the proposed liver (II) and (III), respectively – the same way as in the work of
tumor segmentation method is shown in Figure 1. Lee et al. .
31
2.1. CT image pre-processing module 2 U
W log 1 (II)
Image pre-processing is the preliminary step for most W width
computer vision tasks and is responsible for the final
performance to some extent, especially for medical image 2 W level U
processing tasks. The raw CT images are generally b W log 1 (III)
28
represented in the Hounsfield unit (HU), which is a width
quantitative scale for describing radiodensity. This results W width and W refer to the CT window width and
level
in a wide range of intensity values from −1000 HU to +2000 window level respectively. ε represents the margin between
HU in a typical CT image. However, in clinical practice, the upper and lower bounds and between the window end
only a narrow range is leveraged by radiologists for specific and start gray levels, and it determines the slope of the
tissue types and pathologies. The radiologists adjust a sigmoid function at the center.

Volume 2 Issue 2 (2025) 62 doi: 10.36922/aih.5001