Explainable Model for Localization of Spiculation in Lung Nodules

Abstract

When determining a lung nodule malignancy one must consider the spiculation represented by spike-like structures in the nodule’s boundary. In this paper, we develop a deep learning model based on a VGG16 architecture to locate the presence of spiculation in lung nodules from Computed Tomography images. In order to increase the expert’s confidence in the model output, we apply our novel Riemann-Stieltjes Integrated Gradient-weighted Class Activation Mapping attribution method to visualize areas of the image (spicules). Therefore, the attribution method is applied to the layer of the model that is responsible for the detection of the spiculation features. We show that the first layers of the network are specialized in detecting low-level features such as edges, the last convolutional layer detects the general area occupied by the nodule, and finally, we identify that spiculation structures are detected at an intermediate layer. We use three different metrics to support our findings.