Optimizing spectral power compression with respect to inference performance for recognition of tumor patterns in...

Imaging modalities are widely used to explore and diagnose diseases. Feature extraction methods are used to quantitatively describe and identify objects of interest in acquired images, typically involving data compression. The extracted features are subject to clinical inference, whereby the compression ratio used for feature extraction can affect the inference performance. In this paper, a new method is introduced which allows for optimal data compression with respect to performance maximization of uncertain inference. The model introduced herein identifies objects of interest using selective data compression in the frequency domain. It quantifies the amount of information provided by the inference involving these objects, calculates the inference efficiency, and estimates its cost. By analyzing the effect of data compression on inference efficiency and cost, the method allows for the optimal selection of the compression ratio. The method is applied to prostate cancer diagnosis in ultrasound images.