A hyperbolic evidential learning approach for hierarchical consistency and false-negative minimization in lung carcinoma subtyping

Abstract

Lung carcinoma is among the most prevalent and fatal cancers across the globe, and an accurate clinical classification of the disease subtypes is necessary to plan the treatment.Computed tomography (CT) imaging has become the primary tool in the diagnosis of lung cancer, and automated subtyping with the help of deep learning continues to present significant obstacles. Most of the models available consider subtyping as a flat classification problem and do not consider the hierarchical nature of relationships between the categories of lung carcinoma. Furthermore, these models tend to give confident predictions, but without much indication of reliability of these predictions, thus restricting their application in practical clinical practice. This thesis proposes a hierarchical subtyping of lung carcinoma utilizing hyperbolic evidential learning. The suggested methodology simulates the inherent hierarchy of the lung cancer subtypes in a hyperbolic space, which promotes predictions that are consistent with clinical experiences. Parallel to this, predictive uncertainty is estimated using evidential learning, which enables the model to produce lower confidence when the information in its possession is unclear or insufficient. This combination assists to enhance the consistency as well as reliability of the outputs of the model. The performance of the proposed method is assessed with the use of CT scan dataset to assess the classification performance and the quality of uncertainty estimation. Additionally, four backbone architectures (EfficientNet-B0, DenseNet121, ViT-B/16, and PVT-V2-B0) are evaluated, and Vision Transformer (ViT) performs best among the hyperbolic models, achieving a False Negative Rate of 0.0029 (approximately 99.7 percent sensitivity). The findings suggest that the use of a hierarchical structure and recognizing uncertainty causes more reliable subtype predictions and the improvement of more consistent confidence measures. In general, the present study is expected to encourage more informed clinical decisions and make a contribution to the safe and practical application of artificial intelligence in the diagnosis of lung carcinoma.