CW-CAE: Pulmonary Nodule Detection from Imbalanced Dataset Using Class-Weighted Convolutional Autoencoder

A Class-Weighted Convolutional Autoencoder (CW-CAE) is proposed in this paper to resolve the skewed class distribution found in lung nodule image datasets. The source of these images is the Lung Image Database Consortium image collection (LIDC-IDRI) comprising of lung Computed Tomography (CT) scans. The annotated CT scans are divided into image patches that are labeled as either ‘nodule’ or ‘non-nodule’ images. Understandably, the number of samples containing nodules is substantially less as compared to that of the non-nodules. To solve the class-imbalance issue and prevent bias in decision-making, a class-weight equal to the ratio of the total population to the class population is introduced. The class-weights are multiplied with the respective loss function associated with each class during the computation of the aggregate loss function in the training phase. The training module consists of a feature extractor which is the encoder part of a Convolutional Autoencoder (CAE) pre-trained on the lung nodule dataset, and a classifier comprising of randomly initialized fully connected layers at the output stage. Experiments prove the efficacy of our class-weighted approach for the imbalanced dataset as compared to the state of the art.