Why Registration Quality Matters: Enhancing sCT Synthesis with IMPACT-Based Registration

2025-10-24

AI Summary

This work addresses the problem of registration-induced bias and degraded anatomical fidelity in MRI/CBCT-to-synthetic-CT (sCT) synthesis. To mitigate these issues, we propose an anatomy-aware joint registration-synthesis optimization framework. Our key contributions are: (1) an IMPACT similarity metric—leveraging a pre-trained segmentation network—to enable high-precision rigid/affine registration with improved anatomical consistency; and (2) a 2.5D U-Net++ architecture (with ResNet-34 encoder) trained end-to-end using a hybrid loss combining pixel-wise L1 and IMPACT-Synth perceptual terms, optimized via AdamW, with five-fold ensemble and test-time augmentation. Experiments demonstrate that IMPACT registration significantly reduces MAE and enhances local anatomical plausibility. Although conventional metrics (e.g., Elastix-based) show superior quantitative registration performance on public benchmarks, our analysis reveals their insensitivity to clinically relevant anatomical misalignment—highlighting a critical evaluation bias. Crucially, we empirically validate that high-fidelity anatomical registration is essential for robust sCT generalization.

Technology Category

• Constraint Satisfaction and Optimization: Satisfiability
• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

We participated in the SynthRAD2025 challenge (Tasks 1 and 2) with a unified pipeline for synthetic CT (sCT) generation from MRI and CBCT, implemented using the KonfAI framework. Our model is a 2.5D U-Net++ with a ResNet-34 encoder, trained jointly across anatomical regions and fine-tuned per region. The loss function combined pixel-wise L1 loss with IMPACT-Synth, a perceptual loss derived from SAM and TotalSegmentator to enhance structural fidelity. Training was performed using AdamW (initial learning rate = 0.001, halved every 25k steps) on patch-based, normalized, body-masked inputs (320x320 for MRI, 256x256 for CBCT), with random flipping as the only augmentation. No post-processing was applied. Final predictions leveraged test-time augmentation and five-fold ensembling. The best model was selected based on validation MAE. Two registration strategies were evaluated: (i) Elastix with mutual information, consistent with the challenge pipeline, and (ii) IMPACT, a feature-based similarity metric leveraging pretrained segmentation networks. On the local test sets, IMPACT-based registration achieved more accurate and anatomically consistent alignments than mutual-information-based registration, resulting in improved sCT synthesis with lower MAE and more realistic anatomical structures. On the public validation set, however, models trained with Elastix-aligned data achieved higher scores, reflecting a registration bias favoring alignment strategies consistent with the evaluation pipeline. This highlights how registration errors can propagate into supervised learning, influencing both training and evaluation, and potentially inflating performance metrics at the expense of anatomical fidelity. By promoting anatomically consistent alignment, IMPACT helps mitigate this bias and supports the development of more robust and generalizable sCT synthesis models.

Problem

• Enhancing synthetic CT generation from MRI and CBCT
• Evaluating registration strategies to improve anatomical alignment
• Mitigating registration bias in supervised learning pipelines

Innovation

• 2.5D U-Net++ with ResNet-34 encoder for cross-modality synthesis
• IMPACT-Synth perceptual loss combining SAM and TotalSegmentator
• Feature-based IMPACT registration replacing mutual information metrics