Spatio-temporal deep learning with temporal attention for indeterminate lung nodule classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of cancer death globally.
• Computer-aided diagnosis (CAD) systems improve lung cancer screening accuracy.
• Analyzing serial imaging for indeterminate lung nodules remains underexplored.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for predicting indeterminate lung nodule malignancy using serial screening CT images.
• To enhance the accuracy of malignancy prediction by analyzing the temporal evolution of lung nodules.

Main Methods:

• Introduced a global attention convolutional recurrent neural network (globAttCRNN) integrating spatial and temporal analysis.
• Utilized a 2D CNN for spatial feature extraction and an RNN with a global attention module for temporal feature capture.
• Implemented novel temporal data handling strategies (augmentation, dropout) to address missing data.

Main Results:

• The globAttCRNN achieved an AUC-ROC of 0.954 on an independent test set.
• The model outperformed baseline single-time and multiple-time architectures in malignancy prediction.
• The temporal global attention module effectively prioritized informative time points for nodule analysis.

Conclusions:

• The proposed globAttCRNN demonstrates significant potential for improving the accuracy of indeterminate lung nodule diagnosis.
• This framework can aid radiologists in decision-making and reduce inter-reader variability in lung cancer screening.
• The model offers a valuable tool for lung cancer risk stratification and early detection.