Abstract:
Early detection and precise delineation of liver
tumors are c ritical for improving prognostic outcom es in
patients w ith hepatocellular carc inoma (HCC). This study
introduces two novel deep learning architectures designed for
autom ated liver tumor segm entation and lesion classification
using computed tomography (CT) scans, integrating dom ain
specific data augmentations for enhanced robustness.
The first architec ture, a dual stream convolutional neural
network (CNN) for segmentation, proc esses high resolution
spatial inform ation and broader contextual data in parallel. It
incorporates attention mechanisms to selectively enhance
feature focus and is augmented with random window shifts to
m itigate CT contrast variations. This model achieved a
superior tumor boundary accuracy , demonstrated by a Dice
Similarity Coeffic ient (DSC) of 0.85±0.03, significantly
outperform ing baseline U Net (DSC 0.81±0.04) and 3D U Net
(DSC 0.83±0.03).
The second architecture, a hybrid 3D CNN for lesion
c lassification, leverages full volumetric processing and 3D
channel attention to ac curately distinguish between benign
and m alignant lesions. It demonstrated im proved diagnostic
performance, achieving an accuracy of 0.91±0.03 and an Area
Under the ROC Curve (AUC) of 0.92±0.02, surpassing 2D
baselines such as ResNet (accuracy 0.86±0.04, AUC 0.87±
0.03) and DenseNet (ac curacy 0.88±0.03,
AUC 0.89±0.02).
Valid ated on a diverse m ulti institutional dataset (including
LiTS, 3DIRCADb, and a private cohort), these results confirm
the efficacy of multi stream , attention based, and
augm entation enhanced designs in im proving the reliability
and automation of liver cancer diagnostics, sup porting
clinical d ecision m aking
