Task-based Assessment of Deep Networks for Sinogram Denoising with A Transformer-based Observer

TL;DR

This paper introduces a transformer-based model observer to evaluate deep learning denoising methods in sinogram domain for low-dose CT, effectively approximating traditional observer models and aiding method assessment.

Contribution

A novel transformer-based observer model for sinogram domain evaluation of deep learning denoising methods in low-dose CT is proposed.

Findings

The transformer-based model closely approximates LG-CHO for SKE and BKS tasks.

It effectively assesses CNN-based sinogram denoising methods.

The model demonstrates potential for improving low-dose CT denoising evaluation.

Abstract

A variety of supervise learning methods are available for low-dose CT denoising in the sinogram domain. Traditional model observers are widely employed to evaluate these methods. However, the sinogram domain evaluation remains an open challenge for deep learning-based low-dose CT denoising. Since each lesion in medical CT images corresponds to a narrow sinusoidal strip in sinogram domain, here we proposed a transformer-based model observer to evaluate sinogram domain supervised learning methods. The numerical results indicate that our transformer-based model well-approximates the Laguerre-Gauss channelized Hotelling observer (LG-CHO) for a signal-known-exactly (SKE) and background-known-statistically (BKS) task. The proposed model observer is employed to assess two classic CNN-based sinogram domain denoising methods. The results demonstrate a utility and potential of this transformer-based observer model in developing deep low-dose CT denoising methods in the sinogram domain.