Attention-based Interpretable Deep Learning with Radiomic Features for Pulmonary Nodule Classification
Keywords: Attention, Feature Engineering, Deep Learning, Pulmonary Nodule
TL;DR: We introduce an attention-based interpretable deep learning framework that leverages mathematically predefined, handcrafted features derived from CT imaging for pulmonary nodule classification.
Abstract: Pulmonary nodule classification is critical for early lung cancer screening, enabling timely intervention and evidence-based clinical decision-making. In this study, we introduce an attention-based interpretable deep learning framework that leverages mathematically predefined, handcrafted features derived from CT imaging for pulmonary nodule classification. In contrast to conventional convolutional neural networks (CNNs) that learn complex and often opaque feature representations, our approach prioritizes transparency and reproducibility by using statistically defined intensity features. The architecture is a lightweight multilayer perceptron (MLP) with channel-wise attention. The model was trained on an in-house dataset and validated on two publicly available external datasets: LUNA (n=1,122) and ISBI (n=220), achieving an area under the receiver operating characteristic curve (AUC) of 0.964 (95% confidence interval [CI]: 0.942–0.983) and 0.974 (95% CI: 0.964–0.984), respectively. The integration of channel-wise attention within the MLP architecture enables the model to explicitly learn and assign relative importance to each input feature, supporting feature-level interpretability.
Submission Number: 43
Loading