Semi-Supervised Learning of Visual Attributes for Automated Assessment of Lung Nodule Malignancy

Deep learning models have been successfully applied to lung nodule malignancy analysis but lack interpretability, for example, in linking underlying nodule visual attributes with malignancy prediction. In this paper, we combine visual attribute prediction and malignancy classification tasks for assisted nodule diagnosis while providing imaging characterization. A semi-supervised attribute prediction model is pre-trained first, and then the downstream lung nodule malignancy classification is fine-tuned, maintaining sensitivity to nodule attributes. Using data from 1018 public and 3547 in-house subjects, we demonstrate that the proposed method yields high malignancy classification with a sensitivity of 87.56%, specificity of 81.92%, and an AUC of 92.41%, while also improving model attribute prediction.