Augmented Dual Input CNN (DI-CNN) for the Diagnostic Classification of Lung Nodule Malignancy from CT Scans

Abstract

We demonstrate that Image Augmentation with a Dual-Input CNN architecture can greatly improve the diagnostic classification performance of a model for AI-based lung nodule malignancy classification from CT scan images. Lung cancer is the leading cause of cancer-related death worldwide, but early detection can improve prognosis. Lung cancer screening using Low Dose Computed Tomography (LDCT) has become a standard practice as a way of determining which pulmonary nodules are likely benign and which nodules require biopsy to determine malignancy. In recent years many studies have investigated the use of CNNs for malignancy estimation using LIDC-IDRI. Recent progress has shown that hybrid algorithms that combine CNNs with Radiomic features can achieve a high accuracy. Furthermore, additional studies have shown that Multi-Path CNN (MP-CNN) architectures are more accurate for this task than CNN architectures using a single input path. We present a novel approach using a special case of the MP-CNN in which both inputs are of the same dimensions, for which we coin Dual Input - CNN (DI-CNN). Furthermore, we greatly increase the labeled data volume of the LIDC-IDRI by incorporating rotation-based augmentation. We observe that the DI-CNN is the most accurate version of MP-CNN of several input dimensions in our comparison, thereby demonstrating that it is not necessary for both input paths to take imagery of different dimensions as was previously thought. Furthermore, we show that through data augmentation, it is possible to substantially increase the labeled data volume thereby allowing the DI-CNN algorithm to outperform a state-of-the-art hybrid CNN/Radiomic algorithm for classification of nodule malignancy.