CICRNet: Clinical Information and Category Relation Improve Imbalanced Skin Cancer Diagnosis

Skin cancer is one of the most prevalent cancer types. However, accurate classification of skin cancer is challenging due to the presence of inter-class similarity, intra-class variation and imbalanced data distributions. And most of the state-of-the-art methods usually yield predictions only with l...

Full description

Saved in:
Bibliographic Details
Published in2022 7th International Conference on Signal and Image Processing (ICSIP) pp. 596 - 600
Main Authors Cao, Yunjian, Wang, Hongqiu, Tian, Dan, Wang, Xin, Wu, Shaozhi, Tian, Miao
Format Conference Proceeding
LanguageEnglish
Published IEEE 20.07.2022
Subjects
category-aware module
Clinical diagnosis
Data aggregation
Fuses
Image processing
Learning systems
Lesions
multi-modal data aggregation
Pipelines
skin cancer diagnosis
Online AccessGet full text

Cover

More Information
Summary:Skin cancer is one of the most prevalent cancer types. However, accurate classification of skin cancer is challenging due to the presence of inter-class similarity, intra-class variation and imbalanced data distributions. And most of the state-of-the-art methods usually yield predictions only with lesion images, without taking patient clinical information into account. In this paper, we focus on developing several novel modules to address these issues, including a category-aware (CA) module and a clinical information fusion (CIF) module. Specifically, for the intrinsic issues, we propose a CA module to model the relation of different skin cancer types, thus to exploit more discriminative features. The CA module also adopts a channel attention learning mechanism to suppress noise and emphasize the relevant features. At the end of the classification pipeline, a CIF module that includes an aggregation mechanism, is designed to fuse the features from images and clinical information. By incorporating the CA module and CIF module with a backbone network, the CICRNet is constructed for skin cancer diagnosis. Both the modules can be easily embedded to a wide range of backbone networks and trained efficiently by an end-to-end manner. Comprehensive experiments are conducted on the publicly accessible dataset, PAD-UFES-20. Simulations show that the CICRNet achieves better or comparable results against that of the other state-of-the-art models.
DOI:10.1109/ICSIP55141.2022.9886238