A classification method embedding atypical patterns for distinguishing tumor subtypes in PET/CT images

• Published in: Biomedical signal processing and control 2024-10, Vol.96, p.106663, Article 106663
• Main Authors: Tong, Guoyu, Jiang, Huiyan, Luan, Qiu, Li, Xuena
• Format: Article
• Language: English
• Subjects: Auxiliary diagnosis; Fuzzy classification; Pattern classes; PET/CT; Tumor subtype
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.106663

Cancer is one of the most dangerous diseases worldwide. Accurate cancer subtype classification facilitates both diagnosis and prognosis. Currently, studies focused on cancer subtype classification mainly utilize invasive histopathological images. However, non-invasive PET/CT images are often the first choice for initial screening and diagnosis of tumors. In addition, existing deep learning models strive to find accurate classification boundaries. It is difficult to make a definitive diagnosis with radiological imaging studies alone. To address these issues, we proposed a classification method embedding atypical patterns for distinguishing tumor subtypes in PET/CT images. First, we introduced a novel pattern class division method, including the original classes representing typical patterns and the extended classes representing atypical patterns. We optimized the distribution of feature vectors through a deep metric learning-based method. Then, we proposed a fuzzy classification loss that employs one or more proxies to represent a pattern class. This loss optimizes the spatial distribution of proxies with their corresponding subsets. Finally, we proposed a fuzzy classification method to predict the class of a sample. The experimental results on two datasets show that the proposed model can separate the samples with atypical features and has high accuracy for samples with typical features. For more difficult tasks, the improvement using the proposed model is more obvious. Furthermore, using the proposed model has good noise immunity and interpretability, which is helpful for clinical auxiliary diagnosis. •Develop a PET/CT tumor subtypes classification method embedding atypical patterns.•Introduce a novel pattern class division method.•Propose a fuzzy classification loss that can separate samples with atypical features.•The proposed method can effectively alleviate the label noise.•The proposed method is more consistent with the diagnostic workflow.