Page 67 - AIH-2-2

Artificial Intelligence in Health Improved liver tumor segmentation with dense networks

segmentation from computed tomography (CT) scans is of this framework, cascaded FCNs have become a
vital for examining, diagnosing, and treating liver cancer. common choice in FCN-based work for liver tumor
However, the manual delineation of liver lesions is generally segmentation in recent years. One well-known method is
prone to high inter-rater variability as it depends on the the Cascaded UNets. It combines two UNets to perform
16
experience and knowledge of radiologists and oncologists. segmentation, building on the empirical success of UNet
17
2,3
In addition, the slice-by-slice annotation pipeline from CT in medical image segmentation. To address the limited
18
volumes is time-consuming and tedious. Therefore, the representation power of UNet for this task, cascaded
4-6
development of an automated segmentation system for ResNets employ two deep residual networks to extract
20
19
liver tumors is essential for clinical practice. more discriminative features. However, the quality of the
liver segmentation obtained in the first stage may affect the
However, constructing such a system presents 21
significant challenges for two main reasons. First, the subsequent segmentation of tumors in the second stage,
potentially leading to performance degradation.
shape, size, location, and number of liver lesions exhibit
substantial inter-patient variability. Second, variations in To mitigate this drawback, a three-stage cascaded
7
CT imaging equipment, acquisition protocols, contrast hierarchical convolutional-deconvolutional neural
21
agents, and scanner resolutions result in unpredictable network (HCDNN) introduces an additional stage to
intensity differences between normal liver tissues and refine liver region segmentation. It is worth mentioning
tumors, leading to blurry boundaries, noise, and other that the aforementioned cascaded FCNs-based methods
8,9
artifacts in CT images. are typically optimized by training FCNs sequentially.
22
To further simplify the training phase, Vorontsov et al.
To address this important yet challenging problem,
significant efforts have been made and they can be broadly attempted to train the cascaded FCNs-based network in an
end-to-end manner.
classified into two categories: hand-crafted feature-based
methods and deep learning-based methods. The former Other methods solve this problem from the perspective
23
primarily employs conventional segmentation technologies of data dimensions. For instance, Hu et al. generated
such as intensity thresholding, level set, and region-growing synthetic tumors in CT scans to reduce the annotating
methods. As a simple yet effective algorithm, thresholding computation and enrich small tumor representation.
10
24
intends to classify the foreground and background through Liu et al. proposed a multimodal masked autoencoder
a threshold value. To optimize the threshold selection, for brain tumor segmentation to alleviate the missing
many promising technologies have been proposed, modalities in magnetic resonance imaging (MRI) data,
such as histogram analysis and between-class variance where both random modalities and patches of the remaining
11
maximization. The level set methods are another set of modalities are masked for a reconstruction task. Moreover,
12
classical segmentation algorithms. For example, Smeets the strong performance of large-scale pre-trained models
et al. presented a semi-automatic method to adjust the in downstream tasks has spurred efforts to apply such
13
initial segmentation using a speed function derived from a models in tackling medical segmentation problems. Liu
25
14
pixel classification algorithm, and Hoogi et al. proposed et al. proposed a universal model incorporating text
a novel method to improve level set segmentation through embeddings learned from the contrastive language-image
26
adaptive re-estimation of the local window. Similarly, the pre-training model by Radford et al., thus enhancing the
region-growing algorithm has attracted attention in the domain adaptability across diverse tumor types.
area of liver lesion segmentation, as exemplified by a 3D Despite promising results obtained by the
seeded region-growing algorithm that modeled liver aforementioned methods, their performance could likely
15
tumors using a bag of Gaussians. In addition, numerous be further boosted by addressing challenges related to
machine learning methods leveraging hand-crafted multiscale tumor segmentation and boundary refinement.
features have been developed to segment liver tumors. Cascaded ResNets has provided a feasible strategy to
19
However, given the superiority of a fully convolutional efficiently segment multiscale tumors. Specifically, the input
network (FCN) in medical image segmentation compared image is resized to various scales and the corresponding
to hand-crafted features, this study focuses on deep- multiscale outputs are averaged to produce the final
learning-based methods. segmentation map during the testing phase. However, this
The framework of cascaded FCNs integrates two FCNs straightforward approach incurs high computational cost
for a combined segmentation of the liver and its associated and memory consumption.
tumors; the first FCN identifies the liver region from CT To account for multiscale tumors and further refine
images, whereas the second one segments tumors within tumor boundaries while circumventing those drawbacks,
the identified region. Benefiting from the effectiveness we present an automatic end-to-end framework for liver

Volume 2 Issue 2 (2025) 61 doi: 10.36922/aih.5001