Page 72 - AIH-2-2

Artificial Intelligence in Health Improved liver tumor segmentation with dense networks

results summarized in Table 1 show that the Window Clip 3.3.3. Effectiveness of hybrid loss
method produced a better segmentation result than the No The effectiveness of the hybrid loss was assessed by comparing
Window method. This is because the preset CT window the weighted cross-entropy loss and the hybrid loss. The
increases the grayscale differences among task-specific tumor segmentation results of the test dataset are shown in
tissues in CT images, thereby reducing the learning burden Table 3. Hybrid loss led to a 1% gain in the Dice per case
on the segmentation network and eliminating interference score. Our intuition is that cross-entropy loss, a form of pixel-
from unrelated tissues. Our integrated CT image pre- level supervision, lacks understanding of global information,
processing module achieved the best tumor segmentation leading to unclear edges in the final segmentation. The
result, as it was aimed to find optimal CT window settings. Lovász-Softmax loss is directly oriented toward segmentation

The visualization of some CT images pre-processed evaluation metrics and captures global information, but its
using different techniques is shown in Figure 5. The images training process is unstable. We utilized a dynamic weighting
reveal that the lesion regions were more conspicuous mechanism in the hybrid loss to select the appropriate loss
against the surrounding tissues when our pre-processing function at different stages of model training. To have a
method was used. To have a more quantitative comparison better visual comprehension of the effect of the hybrid loss,
of different pre-processing approaches, we illustrated their we outlined the boundaries of the segmented tumors in
input-output mapping curves in Figure 6. Figure 9. The hybrid loss produced finer tumor boundary
segmentation compared with the cross-entropy loss.
3.3.2. Effectiveness of I -DenseFCN segmentation
2
network To demonstrate the necessity of the combination of two
types of loss functions and the superiority of the dynamic
We investigated the effectiveness of the I -DenseFCN weights, two more experiments concerning the loss were
2
segmentation network by comparing it with the 2D conducted. One is the experiment on the sole Lovász-Softmax
DenseUNet since the I -DenseFCN network is built upon loss, the other is a separate two-stage training scheme, where
2
the 2D DenseUNet by introducing the inter-block dense the network is trained with the weighted cross-entropy
connections in the decoder. Both of the two networks are loss in the first stage and then fine-tuned with the Lovász-
integrated with the proposed CT image pre-processing Softmax loss in the second stage. The results of the above two
module for a fair comparison. The training loss curves supplementary experiments are shown in Table 3. The results
of the two networks are depicted in Figure 7. The show that a joint utilization of the two types of loss functions
I -DenseFCN network achieved a lower training loss value achieved better tumor segmentation performance than either
2
compared to the 2D DenseUNet. We argue that the skip of them alone. On the other hand, the design of dynamic
connections between modules achieve multiscale feature weights outperformed the two-stage training strategy.
fusion, which helps to enhance the spatial information
of different scales in the output feature maps. This spatial 3.4. Comparison with other methods on LiTS
information promotes the recognition of small-scale Since the proposed model is designed based on 2D FCNs
tumors. The Dice per case scores of tumors (Table 2) due to the limitation of computer resources, its tumor
indicate that the proposed segmentation network produces segmentation results were compared against published
better segmentation results than the 2D DenseUNet, 2D-based methods (Table 4). Most methods employed
demonstrating the effectiveness of inter-block dense FCNs but with some variations. Some methods attempted
connections of the I -DenseFCN network.
2
To visualize the effectiveness of the proposed Table 2. Comparison of different segmentation networks
segmentation network, we present the segmentation
results of some samples from the validation dataset Segmentation network Dice per case (%)
in Figure 8. The I -DenseFCN network handled the 2D DenseUNet 69.0
2
segmentation of multiscale tumors more effectively than I -DenseFCN 69.6
2
the 2D DenseUNet, as expected.
Table 3. Comparison of different loss functions (Dice: %)
Table 1. Comparison of different pre-processing methods Loss function Dice per case (%)

Pre-processing method Dice per case (%) Lovász-Softmax 65.5
No window 67.0 Weighted cross-entropy 69.6
Window clip 68.7 Two-stage training 70.1
Window optimization 69.0 Hybrid loss 70.6

Volume 2 Issue 2 (2025) 66 doi: 10.36922/aih.5001