Answers are grounded in this paper and cite their sources.
The limited interpretability of deep neural networks hinders their implementation in clinical practice, in particular in the analysis of computed tomography (CT) of the lungs. Therefore, an important problem is to develop decision support systems capable of generating interpretable descriptions of the characteristics of pulmonary nodules and assessing the risk of their malignancy. In this study, a decision support system for explainable automatic characterization of lung nodules in CT images and their malignancy assessment has been presented. A modified architecture of the neural additive model, which makes it possible to represent all parameters of the target distribution as additive nonlinear functions of the input features, ensuring interpretability at the feature level, has been introduced. The proposed system includes a chain of modules for lung nodule segmentation, lesion feature extraction, and feature contribution explanation. The model has been trained using the publicly available LIDC-IDRI dataset and has achieved accuracy comparable to the baseline neural additive model architecture, while providing interpretable, clinically relevant results and additional explanatory power. Based on an analysis of the obtained shape functions, it has been shown that the results are consistent with existing data in the field of radiology. The developed system has provided automatic generation of structured descriptions of lesions and visualization of the contribution of each feature to the assessment of malignancy and dispersion.
No takes yet. Share an insight, caveat, or question.
Dumaev et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: