Page 71 - AIH-2-2

Artificial Intelligence in Health Improved liver tumor segmentation with dense networks

was proposed for multiclass semantic segmentation, employed to evaluate our method for 3DIRCADb dataset.
16
aiming for direct optimization of the mean intersection- In the training phase, three adjacent slices of size 224 × 224
over-union (mIoU) score. It shows substantially better were randomly cropped from raw CT volumes to train our
segmentation performance than the cross-entropy loss on I -DenseFCN. No other pre-processing operations were
2
several semantic segmentation datasets. performed in the period.
The cross-entropy loss is unable to take the regional Similar to the 2017 LiTS challenge (XX), the main
36
information into account during the optimization process, evaluation metric to assess the performance of liver tumor
but it is stable. Despite the improved segmentation segmentation was the Dice per case score, which refers to
performance achieved by the Lovász-Softmax loss, it the average Dice score per volume.
is unstable in the training process. Hence, the Lovász-
Softmax loss is exploited in a fine-tuning stage to refine 3.2. Implementation details
27
the segmentation results in the original study. The I -DenseFCN model was implemented under the
2
37
In this work, a hybrid loss is designed by combining PyTorch framework and was trained with Adam
38
the cross-entropy loss and the Lovász-Softmax loss with optimizer. The initial learning rate was set to 0.001 and
dynamic weights. The weights of two types of losses are decayed every 40 epochs at a rate of 0.1. The training of each
linearly determined by the iteration of the training process. model required approximately 30 h on a single NVIDIA
The hybrid loss function is defined in Equation (V): GTX 1080Ti GPU for 120 epochs, with a minibatch size
of 16. The inference time cost on the LiTS test dataset was
n n about 90 s/case.
L 1 L wce L ls (V)

N N In the training stage, a liver segmentation model based
on the 2D DenseUNet was first trained with CT slices in
29
Where L and L represent the weighted cross-
wce
ls
entropy loss and the Lovász-Softmax loss, respectively. their original size. Then the proposed model was trained
N is the total number of training iterations and n is the on the cropped dataset. To alleviate overfitting, standard
index of the training iteration. α is a hyperparameter data augmentation techniques, such as random flipping,
chosen to balance L and L magnitude and is rotating, and scaling, were applied. In the test stage, for a
wce
ls
determined empirically. The hybrid loss puts more weight given 3D CT volume, a coarse liver segmentation result
on the cross-entropy loss in the early stage of the training was first generated, and then the liver bounding box was
process and shifts gradually to the Lovász-Softmax loss used by the proposed model to predict the final liver tumor
in the late stage. This leads to a coarse-to-fine liver tumor segmentation.
segmentation process. 3.3. Analysis of the proposed method
3. Experiments In this section, we conducted several experiments on
the LiTS dataset to evaluate the contributions of each
3.1. Datasets and evaluation metrics
component in our proposed I -DenseFCN, including
2
Experiments were conducted to evaluate the performance the CT image pre-processing module, the I -DenseFCN
2
2
of the proposed I -DenseFCN using two public datasets, segmentation network, and the hybrid loss.
namely 2017 LiTS challenge and 3DIRCADb. The LiTS
36
dataset contains 201 contrast-enhanced abdominal CT 3.3.1. Effectiveness of CT image pre-processing
scans collected from multiple clinical sites, incorporating module
large variations in the in-plane resolution and the slice To validate the effectiveness of the CT image pre-
spacing due to the difference in scanners and protocols; processing module, we compared the tumor segmentation
the dataset includes 131 cases for training and 70 cases for performance of different CT image pre-processing
testing. The axial slices of all scans have an identical size approaches using the 2D DenseUNet model. The
of 512 × 512, and the number of slices in each case ranges approaches included HU value clipping with a preset
from 42 to 1026. The 3DIRCADb dataset is composed CT window (Window Clip), the full range of HU values
of 20 3D CT scans, of which 15 volumes have hepatic without any window (No Window), and the proposed
tumors in the liver. The number of slices in each case varies integrated window optimization pre-processing module
between 74 and 225. (Window Optimization). The preset CT window was set
In our experiments, 131 training cases were further split to [−200, 250]. The initialization parameters W width, W level,
into a training set and validation set at a ratio of 111:20 U, and ε for window optimization pre-processing module
for LiTS dataset, whereas 2-fold cross-validation was were set to 450, 50, 255, and 1, respectively. Experimental

Volume 2 Issue 2 (2025) 65 doi: 10.36922/aih.5001