Task-Adaptive Low-Dose CT Reconstruction

TL;DR

The paper tackles the problem that low-dose CT reconstructions, while scoring well on traditional image fidelity metrics, often fail to preserve information essential for diagnostic tasks. It introduces a task-adaptive reconstruction framework that freezes a pre-trained task network and uses its outputs as a regularizer in the reconstruction loss, balanced by a weight $\alpha$, to preserve task-relevant features without diverging from plausible reconstructions. Applied to liver and liver tumor segmentation, the method achieves Dice scores near full-dose performance (around $0.706$–$0.707$) and significantly outperforms joint-training baselines and traditional methods. This approach provides an easily integrable path to produce diagnostically faithful reconstructions from LDCT, with potential to improve clinical adoption and task-driven quality assessment in CT imaging.

Abstract

Deep learning-based low-dose computed tomography reconstruction methods already achieve high performance on standard image quality metrics like peak signal-to-noise ratio and structural similarity index measure. Yet, they frequently fail to preserve the critical anatomical details needed for diagnostic tasks. This fundamental limitation hinders their clinical applicability despite their high metric scores. We propose a novel task-adaptive reconstruction framework that addresses this gap by incorporating a frozen pre-trained task network as a regularization term in the reconstruction loss function. Unlike existing joint-training approaches that simultaneously optimize both reconstruction and task networks, and risk diverging from satisfactory reconstructions, our method leverages a pre-trained task model to guide reconstruction training while still maintaining diagnostic quality. We validate our framework on a liver and liver tumor segmentation task. Our task-adaptive models achieve Dice scores up to 0.707, approaching the performance of full-dose scans (0.874), and substantially outperforming joint-training approaches (0.331) and traditional reconstruction methods (0.626). Critically, our framework can be integrated into any existing deep learning-based reconstruction model through simple loss function modification, enabling widespread adoption for task-adaptive optimization in clinical practice. Our codes are available at: https://github.com/itu-biai/task_adaptive_ct

Figures (4)

• Figure 1: Overview of the proposed task-adaptive reconstruction framework with data preprocessing and benchmarking. Full-dose scans and ground truth task outputs are taken from a source dataset. Corresponding low-dose scans are simulated following Leuschner et al. lodopab-ct. Low-dose scans are first processed by the task-adaptive reconstruction network to produce reconstructed images. The reconstructions are compared against the ground truth full-dose scans via the reconstruction loss. Simultaneously, the reconstructed images are fed into a frozen pre-trained task network to produce task predictions, which are compared against ground truth task outputs via the task loss. The two losses are combined using a weight hyperparameter, where the pre-trained task network acts as a regularizer to guide the training to preserve diagnostically relevant features. Reconstructed images and ground truth full-dose scans are compared to calculate PSNR and SSIM scores. Predicted and ground truth task outputs are compared to calculate the task-specific performance metric.
• Figure 2: A sample from the preprocessed dataset showing simulated low-dose scan, original full-dose scan and the corresponding segmentation map. Purple area shows the background, cyan area shows the liver and yellow area shows the liver tumor.
• Figure 3: A sample with low-dose scan, its corresponding reconstructions and ground truth full dose scan. PSNR, SSIM and Dice scores are given above the visuals.
• Figure 4: Predicted segmentation maps of pre-trained segmentation model for the sample in Fig. \ref{['reconstructions-figure']} from the low-dose scan, its corresponding reconstructions and ground truth full dose scan. Also, ground truth segmentation map is provided for reference. PSNR, SSIM and Dice scores are given above the visuals.