Purpose: Liver cancer is one of the leading causes of cancer death. Accurate detection of liver cancer early using CT could assist doctors in disease diagnosis and treatment planning. This study proposes a deep learning method for accurate detection and segmentation for liver tumors, especially for small liver tumors.
Methods: This study included 131 CTs of patients with liver cancer. We hypothesize that liver tumors of different sizes share similar imaging characteristics. The information of known tumors can compensate for the information loss of small tumors from the feature propagation, which can improve small tumor segmentation. Our method first constructs a liver tumor template by all CT slices containing tumors in the training set, then extracts the semantic features of input CT image and tumor template simultaneously by using two residual networks. Finally, the relationship between input CT and tumor template in the feature space is exploited to improve liver cancer segmentation.
Results: Among 20,693 CT slices of the 31 testing patients, all CT slices were separated into groups according to tumor size as follows: 0.1–2.0 (3.17%), 2.1–5.0 (3.58%), 5.1–10.0 (3.13%), and 10.1–20.0 (3.01%) cm. Other slices without tumor or tumor size > 20 cm were categorized as another same group. Our method outperforms state-of-the-art models including Unet, PAN, DeepLabV3, FPN, LinkNet, and PSPNet on different sizes of tumors, especially for small liver tumor segmentation. For the 10.1-20cm liver tumor, our method achieved 7.1%, 1.9%, 2.9%, 3.8%, 3.3%, 1.3% improvement, and on the 0.1-2.0cm small liver tumor set, our method achieved 8.4%, 10.0%, 11.3%, 9.1%, 10.9%, and 9.6% improvement, respectively.
Conclusion: This work indicates tumors of different sizes share similar imaging characteristics. The small-large tumors relation can significantly improve small liver tumor segmentation, which is beneficial for disease diagnosis and treatment planning.
Not Applicable / None Entered.