Distribution-Aware Loss for Lesions Detection Using White-Light Endoscopy in Colorectal Region

• Published in: IEEE access Vol. 12; pp. 46138 - 46148
• Main Authors: Sun, Wei, Zhao, Ruiqi, Zhang, Kunpeng, Gao, Junbo, Qu, Guoqiang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Biomedical imaging; Classification; Colorectal cancer; Complexity theory; Deep learning; Diagnostic systems; Endoscopes; Endoscopy; Lesions; Medical diagnostic imaging; medical image process; Medical imaging; Predictive models; Task analysis; Training; White light; white-light endoscopy
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3381614

Computer-aided diagnostic systems have evolved into critical tools for endoscopists in diagnosing and reducing missed diagnoses. However, due to the lower incidence of rare diseases in medical images compared to common diseases, there exists an imbalance in sample distribution for lesion classification. This imbalance results in reduced detection and classification accuracy. In the realm of deep learning, detection accuracy is influenced not only by the model but also by the choice of the loss function. This study introduces a novel solution to address the imbalance issue of colorectal lesions in white light endoscopy by proposing a loss function named Label Distribution and Scale Distribution Aware Loss (LSDA-Loss). Our innovative loss function resolves the category imbalance problem by considering sample distribution and employing Bayesian equations to quantify the degree of imbalance. Furthermore, we adopt proportional distribution to evaluate the complexity of categorizing each sample. Experimental results from three independent datasets demonstrate that: 1) the integration of the proposed loss function with three typical FPN models significantly enhances detection accuracy, achieving improvements of up to 94.56%. 2) Our loss function effectively balances detection accuracies across the three categories, surpassing the performance of the original loss function.