Deep Learning Approach to Temporal Dimensionality Reduction of Volumetric Computed Tomography

A common approach for analyzing medical images on volumetric data is to employ deep 2D convolutional neural networks (2D CNN) on each individual slice of the volume. This is largely attributed to the challenges posed by the nature of three-dimensional data: variable volume size, high GPU and RAM requirements, costly parameter optimization, etc. However, handling the individual slices independently in 2D CNNs deliberately discards the temporal information that is contained within the depth of the volumes, which tends to result in poor performance. In order to maintain temporal information, current solutions reduce the temporal dimensionality of the data using non-adaptive sampling processes, such as taking slices at given intervals or by means of interpolation. However, although this allows to keep some temporal information, it may discard important slices. In this paper, we propose a method based on GradCam to select meaningful slices in computed tomography volumes by evaluating the activation map to reduce temporal data dimensionality. Extensive experiments demonstrate the effectiveness of our method when compared to the current state of the art.