Related Experiment Video
Updated: Aug 17, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
ASE-Net: A tumor segmentation method based on image pseudo enhancement and adaptive-scale attention supervision
Junzhi Zhang1, Huiyan Jiang2, Tianyu Shi1
1Software College, Northeastern University, No. 195, Chuangxin Road, Hunnan District, Shenyang, 110169, Liaoning, China.
Computers in Biology and Medicine
|December 14, 2022
Summary
This study introduces ASE-Net, a new model for segmenting tumors in PET/CT scans. It improves accuracy by generating pseudo-enhanced CT images and using an adaptive-scale attention module for better tumor detection across various sizes.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Fluorine 18(18F) fluorodeoxyglucose positron emission tomography and Computed Tomography (PET/CT) is crucial for cancer diagnosis and treatment.
- Accurate tumor segmentation in PET/CT is challenging due to low contrast and varying tumor scales.
Purpose of the Study:
- To develop a novel multimodality tumor segmentation model, ASE-Net, to overcome limitations in current PET/CT imaging analysis.
- To enhance the accuracy and discriminability of tumor segmentation by integrating metabolic and structural information.
Main Methods:
- Proposed a pseudo-enhanced CT image generation method using metabolic intensity to improve spatial learning and structural discriminability.
- Introduced an Adaptive-Scale Attention Supervision Module within skip connections to provide scale-specific receptive fields for tumors.
- Utilized a Dual Path Block as the network backbone, incorporating residual learning and dense connections for effective feature extraction.
Main Results:
- ASE-Net achieved high performance on two clinical PET/CT datasets, with Dice Similarity Coefficients of 78.56% and 72.57%.
- The model demonstrated superior performance compared to state-of-the-art methods in segmenting both large and small tumors.
- The proposed approach effectively addresses challenges posed by low-contrast imaging and diverse tumor scales.
Conclusions:
- ASE-Net offers a significant advancement in multimodality tumor segmentation for PET/CT imaging.
- The model's ability to handle varying tumor scales and improve image discriminability can aid pathologists in making more accurate diagnoses.
- Improved diagnostic accuracy through advanced imaging analysis has the potential to enhance patient survival rates.
