Reducing False-Positive Detections Using the Distance Between Activation Distributions in Individual Channels

We propose a novel method to estimate the confidence of outputted predictions of a convolutional neural network. We show that different channels in one layer can be treated as an ensemble and extract the confidence of a prediction from a single channel. To achieve this, we compute statistical distances between activation distributions located at the predicted mask and its surrounding area and aggregate it across all channels in a deep layer of a network. Research on a segmentation network of lung cancer nodules from 3d computer tomography images has shown growth of precision compared to the thresholding output network values. The more layers used to compute confidence, the better performance obtained, allowing for up to 18% fewer false-positives detections on the source Cancer dataset and up to 54% fewer false-positives detections on an unseen Covid dataset. Analyzing channel activations doesn’t require any changes in the training procedure with a negligible amount of additional computations at the inference time.