Design and Evaluation of Deep Learning Based Inverse Filter for a Computer-Aided Hot Spot Detection Support System of Bone Scintigrams

Abstract

This paper proposes a deep learning-based method to design inverse filters for image processing filters in bone scintigrams with undisclosed specifications. Although image processing filters remove noise and generate images suitable for medical professional interpretation, they can degrade the performance of a hot spot detection support system trained with non-filtered images. The proposed method designs an inverse filter that predicts the image before the filter is applied from the filtered image. This paper presents the results of applying the proposed method to images processed with two filter types from GE and Siemens Healthcare, and experimentally demonstrates that non-filtered images can be predicted with good accuracy. It is also confirmed that predicted non-filtered images improve the accuracy of a hot spot detection support system trained without filtered images.