.png){kind=icon}
.png){kind=icon}
Return
Tumor classification algorithm via parallel collaborative optimization of single- and multi-objective consistency on PET/CT
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asoc.2024.112245.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Malignant tumors still have a high incidence and mortality rate worldwide. Pathological examination remains the clinical gold standard for tumor diagnosis. However, some patients cannot undergo pathological examination due to advanced age and special lesion location. Therefore, making full use of PET/CT to assist doctors in tumor classification has important clinical significance. Since category labels are calibrated according to pathological images, it is difficult to obtain effective pathological category features directly using PET-CT image modeling. In response to this problem, this paper proposes a novel tumor classification algorithm. This method fully utilizes multi-gray-level 3D gray-level co-occurrence matrix and the proposed rough and fine constraint network under the constraint loss of rough and fine labels. Based on single- and multi-objective consistency, a parallel collaborative optimization method is proposed, including category consistency loss and feature specificity loss. To reduce the interference of redundant features, an improved Boruta feature selection method using multiple classifiers and multiple steps is proposed. The final result is obtained through a conditional weighted voting function. The proposed method shows excellent performance in both the submodels and the fusion model. We validated the proposed tumor classification method on three datasets and achieved good performance with the accuracy of 0.80-0.85 and F1-score of 0.78-0.88. The results indicate that the proposed method has good performance and generalization ability.
Keywords:
Gray level co-occurrence matrix
Multi-label classification
Loss function
Feature optimization
Journal
IF:
6.6
Papers:
1.4W
Citations:
4.8W