Enhancing signal detectability in learning-based CT reconstruction with a model observer inspired loss function

TL;DR

This work addresses the mismatch between pixel-wise losses and task-based signal detection in sparse-view CT reconstruction. It introduces a model-observer–inspired signal promoter loss that promotes preservation of weak signals by encouraging reconstructions to reveal planted signals, and combines it with standard MSE training on a U-Net. Across simulated breast CT data, the approach improves signal detectability (AUC) with only modest increases in MSE, outperforming simple add-back strategies and showing robustness to training-signal placement variations. The study also investigates parameter selection via a weaker observer and discusses broader implications for task-driven learning in CT, along with limitations and future directions for real-data validation and architectural generality.

Abstract

Deep neural networks used for reconstructing sparse-view CT data are typically trained by minimizing a pixel-wise mean-squared error or similar loss function over a set of training images. However, networks trained with such pixel-wise losses are prone to wipe out small, low-contrast features that are critical for screening and diagnosis. To remedy this issue, we introduce a novel training loss inspired by the model observer framework to enhance the detectability of weak signals in the reconstructions. We evaluate our approach on the reconstruction of synthetic sparse-view breast CT data, and demonstrate an improvement in signal detectability with the proposed loss.

Figures (8)

• Figure 1: A CNN trained to perform sparse-view CT reconstruction by minimizing a pixel-wise mean-squared-error (MSE) loss is prone to wiping-out weak "signals" in a breast CT phantom. Left panel shows the ground truth phantom with a contrast-detail insert. Middle panel shows the simulated noisy, sparse-view FBP reconstruction. Right panel shows the reconstruction with a CNN trained with a pixel-wise MSE loss taking the noisy FBP image as input. Note many of the weak signals in the contrast-detail insert are clearly visible in the FBP image, but missing from the CNN reconstruction. More information on the data, network architectures, and training protocol used to generate this figure is given in \ref{['sec:methods']}. All images are shown on the scale $[0.174, 0.253] \text{cm}^{-1}$.
• Figure 2: Examples of training pairs used in this study. Top row: simulated noisy 128-view FBP images. Bottom row: corresponding ground truth phantoms. All images are shown on the scale $[0.174, 0.253] \text{cm}^{-1}$.
• Figure 3: Illustration of signal model used in this study: (left) ideal Gaussian signal in object domain, (right) its counterpart in image space discretized to a $512\times 512$ pixel resolution. Both images are cropped to a $7~\textrm{mm} \times 7~\textrm{mm}$ region centered at the signal, which corresponds to a $20\times 20$ pixel region in image space.
• Figure 4: Reconstructions of a test phantom using CNN trained with the proposed signal promoter loss. Far left shows the reconstruction obtained from a CNN trained with MSE loss only, and with increasing weighting factor $\lambda$ on the signal promoter loss to the right. Note that more fine detail and noise becomes visible in the reconstructions as $\lambda$ increases. All images are shown on the scale $[0.174, 0.253] \text{cm}^{-1}$.
• Figure 5: Illustration of signal-known-exactly/background-known-exactly task used in assessing signal detectability performance of reconstruction approaches. All images are shown on the scale $[0.174, 0.253] \text{cm}^{-1}$.