Page 70 - AIH-2-2

Artificial Intelligence in Health Improved liver tumor segmentation with dense networks

Figure 3. The segmentation network architecture. Figure created by the authors.
Abbreviation: Conv: convolutional layer.

task – we introduce dense connections between
upsampling blocks in the decoder, following the approach
34
of Bilinski and Prisacariu. These inter-block dense
connections allow the current upsampling block to use
semantic feature maps from all preceding upsampling
blocks at different scales, thereby transforming the high-
level semantic information into a spatial resolution that
matches the original input image. Since the features
are fused by element-wise summation operation, they
should have the same spatial resolution and number of
channels. In view of this, 3 × 3 convolutions are applied
to the inter-block dense connection paths to adjust the
number of channels, followed by bilinear interpolations
to upscale the spatial resolution. Intra-block and inter-
block dense connections allow for efficient information
propagation among layers and blocks. Furthermore, Figure 4. Dense connections between two upsampling blocks
U-Net-like long skip connections are added between the Abbreviation: Conv-BN-ReLU: Convolutional layer followed by Batch
encoder and decoder to combine high-level semantic Normalization and the Rectified Linear Unit activation function.
features and low-level detail information. The detailed
architecture between two upsampling blocks is shown in where ˘ y denotes the probability of predicting voxel i as
c
Figure 4. i
c
class c (background, liver, or lesion), y denotes the binary
i
c
2.3. Hybrid loss function with dynamic weights label indicating whether voxel i belongs to class c, and ω
i
denotes the weight assigned to class c.
Deep learning-based segmentation frameworks rely not
only on the design of network architecture but also on The cross-entropy loss treats each pixel as an isolated
the choice of loss function. Liver tumor segmentation is sample when calculating the overall loss; it does not
essentially a pixel-level classification task, so the cross- directly correlate with segmentation quality as it is a pixel-
entropy loss is a natural choice for loss function. The wise loss. However, most of the segmentation metrics are
weighted cross-entropy loss as (IV) is employed in practice evaluated based on the overlap region. There are some trials
to tackle the strong imbalance issue in the liver tumor to replace the cross-entropy loss on the medical image
segmentation task. segmentation tasks, such as Dice loss and Jaccard loss .
35
21
Both of them aim to directly optimize the overlapping
1

L y y, ˘ N i1 3 c1 Wy log y c ^ i (IV) between the predicted and true segmentation regions.
C
N
C
i
wce
i
27
Recently, a new kind of loss named Lovász-Softmax loss
Volume 2 Issue 2 (2025) 64 doi: 10.36922/aih.5001