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

Megan Lantz, Emil Y. Sidky, Ingrid S. Reiser, Xiaochuan Pan, Gregory OngiearXiv 2024 · DOI · arXiv

Abstract

Standard deep learning models for CT reconstruction often prioritize noise reduction at the expense of low-contrast feature preservation. Training with Mean Squared Error (MSE) tends to "over-smooth" images, potentially erasing critical diagnostic signals such as microcalcifications. This work introduces a "Signal Promoter" (SigPro) loss function, inspired by model observer frameworks in medical physics. By explicitly penalizing the loss of signal detectability, the proposed method significantly improves the visibility of subtle features. In simulated breast CT experiments, the SigPro loss increased the Area Under the Curve (AUC) for signal detection from 0.53 (MSE baseline) to 0.83, demonstrating a substantial boost in diagnostic utility with minimal impact on overall image fidelity.

Introduction

In medical imaging, the primary goal of reconstruction is to facilitate accurate diagnosis rather than simply producing visually pleasing images. However, standard deep learning approaches optimized for pixel-wise fidelity (e.g., MSE) often act as low-pass filters. In the context of breast CT, this leads to the erasure of tiny, low-contrast features like microcalcifications, which the network misidentifies as noise. This study addresses this misalignment by introducing a task-specific loss function that prioritizes signal detectability.

Methods

The authors propose a hybrid loss function: $L = MSE + \lambda \cdot SigPro$. The SigPro term is derived from the "model observer" framework, measuring the correlation between a known true signal and the network's reconstruction of that signal.

  1. Architecture: A modified U-Net with residual skip connections is employed to facilitate the flow of high-frequency information.
  2. Strategy: The model is first pre-trained on MSE to establish a stable baseline and then fine-tuned with the SigPro term to enhance feature preservation. Randomizing signal locations during training prevents the network from over-sensitizing to specific regions.

Datasets

The study utilized a stochastic breast model to generate 1,000 pairs of noisy, sparse-view Filtered Back-Projection (FBP) images and their noise-free ground truths.

Training data examples showing noisy 128-view FBP inputs (top) and corresponding ground truth phantoms (bottom).

Signals were modeled as low-contrast Gaussian peaks (0.44 mm width, 0.04 cm⁻¹ contrast) designed to mimic the difficulty of finding early-stage calcifications in clinical practice.

Key Results

Performance was quantified using the Area Under the Curve (AUC) of a Channelized Hotelling Observer (CHO)—a standard mathematical proxy for human observer performance.

  • Quantitative Improvement: The MSE-only baseline achieved an AUC of 0.53, indicating poor detectability (barely better than a random guess). The addition of the SigPro loss improved the AUC to approximately 0.83.
  • Visual Evidence: Qualitative analysis using contrast-detail inserts confirmed these findings. As shown in Figure 2, the MSE-trained network significantly blurs or entirely removes the grid of small dots that are clearly visible in the ground truth.
Visual comparison showing how standard MSE loss fails to recover low-contrast signals (dots) present in the ground truth.

While the MSE-only version (right) appears "cleaner" than the noisy FBP (middle), it fails the diagnostic task by erasing the signal grid. The authors report that SigPro-trained networks successfully recover these features, achieving higher detectability at similar levels of reconstruction error.

Discussion and Future Work

The SigPro loss function provides a robust mechanism for balancing noise reduction with the preservation of diagnostic signals. However, the study currently relies on simulated data, and there is no automatic method for selecting the optimal weighting parameter $\lambda$. If $\lambda$ is too high, the network may introduce spurious artifacts. Future work will focus on validating this approach on clinical datasets and exploring unsupervised adaptations where ground truth images are unavailable.

Summary updated

BibTeX
@misc{lantz2024enhancing,
  author = {Megan Lantz and Emil Y. Sidky and Ingrid S. Reiser and Xiaochuan Pan and Gregory Ongie},
  title = {{Enhancing signal detectability in learning-based CT reconstruction with a model observer inspired loss function}},
  year = {2024},
  eprint = {2402.10010},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2402.10010},
}
Figure