The Problem of Inter-Patient CT Registration In lung cancer radiotherapy research, it is valuable to compare CT scan data across different patients to understand how radiation dose distribution affects outcomes. However, each patient's anatomy is unique - lungs differ in size, shape, and position; tumors appear at different locations. Aligning (registering) CT scans from different patients is technically very challenging, especially when tumors are present.
Why Tumors Make Registration Hard Standard image registration algorithms try to make all structures match between two scans. But when one patient has a large tumor where another patient has healthy lung tissue, naively matching the images creates physically impossible distortions - the algorithm incorrectly deforms the lung tissue to fill the space occupied by the other patient's tumor. This ruins the accuracy of downstream analyses.
Voxel-Based Analysis (VBA) Applications VBA is a population-level analysis technique that maps each point (voxel) of every patient's scan to a common coordinate space, then tests whether radiation dose at specific anatomical locations correlates with treatment outcomes. VBA requires high-quality inter-patient registration to ensure that the same anatomical location is compared across all patients.
TRACER Solution TRACER (Tumor-Aware Recurrent Inter-patient deformable image REgistration) is a deep learning method that explicitly accounts for tumor presence during registration. It uses a recurrent neural network architecture (3D Convolutional LSTM) that progressively refines alignment while applying bidirectional tumor rigidity constraints to preserve tumor geometry and prevent unrealistic deformations.
3D Convolutional LSTM Architecture TRACER uses a 3D Convolutional Long Short-Term Memory (3D-CLSTM) encoder-decoder network. Unlike single-step registration methods, the CLSTM registers images progressively across multiple time steps, with each step refining the previous alignment. This recurrent approach captures deformations at varying scales - coarse global alignment first, then fine local refinements.
Bidirectional Tumor Rigidity Loss The key innovation is a bidirectional tumor conditioning approach with two complementary loss terms: the 'tumor preservation loss' constrains the moving patient's tumor voxels to deform rigidly (preserving tumor volume and shape); the 'tumor obliteration loss' constrains the fixed image's tumor region to deform to healthy tissue when aligned with the moving patient. Together, these prevent the twin failure modes of tumor erosion and unrealistic distortion.
Why Bidirectional Matters In inter-patient registration, tumors can appear in either the 'moving' image (being warped) or the 'fixed' image (reference). Without bidirectional treatment: if the tumor is in the moving image, it gets deformed away to match healthy tissue in the fixed image; if the tumor is in the fixed image, the algorithm creates unrealistic distortions in the moving image around the tumor location. TRACER handles both cases explicitly.
Unsupervised Training TRACER was trained unsupervised (without ground-truth deformation fields) on 32,220 CT image pairs from 180 patients with locally advanced NSCLC. This is a critical advantage because ground-truth inter-patient deformation fields cannot be measured in practice - they require careful unsupervised learning that relies only on image similarity metrics and the anatomical constraint losses.
Three Independent Test Datasets TRACER was evaluated on three completely independent test datasets: Dataset I (380 image pairs from 20 patients, public dataset); Dataset II (756 image pairs from 28 patients, public dataset with diverse tumor sizes from 0.022 cc to 640 cc across all lung lobes); and Dataset III (42 institutional LA-NSCLC patients treated with definitive IMRT to 60 Gy, with radiation dose maps available for dose accuracy assessment).
Tumor Preservation Metrics Registration accuracy was assessed with multiple tumor-specific metrics: tumor volume loss (the percentage change in tumor voxel count after deformation), local expansion and shrinkage inside the tumor (using Jacobian determinant analysis), and mean squared error between original and deformed tumor masks. Smaller values indicate better tumor geometry preservation.
Normal Tissue Alignment Metrics For healthy tissue alignment, Dice Similarity Coefficient (DSC) and 95th-percentile Hausdorff distance (HD95) for lungs, heart, and spinal cord were measured. Tubular organ alignment was assessed by Median of Closest Points Distance (MCD) for the pulmonary artery, aorta, inferior vena cava, and trachea.
Radiotherapy Dose Accuracy In Dataset III, the planned tumor dose difference (delta-PTD) measured how well the radiation dose map was preserved after registration. This is the most clinically relevant metric for VBA applications in radiotherapy - accurately transferring dose maps between patients is essential for identifying dose-outcome correlations.
Best Tumor Preservation Across All Datasets TRACER achieved the lowest tumor volume loss and local shrinkage metrics across all three test datasets. The mean tumor volume difference was only 0.24%, dramatically better than PACS (which showed complete tumor erosion in some cases) and better than FastSym and Transmorph. This confirms that tumor conditioning is critical - without it, registration algorithms aggressively deform tumor voxels to match the healthy tissue in the other patient.
Second Best for Normal Tissue Alignment While TRACER excelled at tumor preservation, it ranked second to PACS for normal tissue and tubular organ alignment. PACS achieved better pulmonary artery, aorta, and vena cava alignment. However, PACS completely failed at tumor preservation, representing an unacceptable trade-off for tumor-containing CT registration.
Lowest Radiation Dose Error TRACER achieved the smallest planned tumor dose error in Dataset III, directly demonstrating its utility for radiotherapy outcome research. PACS, despite good normal tissue alignment, had the largest dose error due to tumor geometry distortion. This confirms that preserving tumor geometry is the most important property for VBA-based radiotherapy dose-outcome studies.
Fewer Excluded Patients TRACER excluded fewer patients due to poor registration than the commonly used iterative SyN method, which excluded 45% or more of patients. This is practically important because high exclusion rates reduce statistical power in VBA population studies, potentially causing false negative findings.
Tumor Conditioning is Critical Ablation experiments removing tumor conditioning showed dramatic degradation in tumor preservation accuracy - the model aggressively deformed tumor voxels to align with healthy tissue in the other patient, exactly the problem TRACER was designed to solve. This confirms that explicit tumor conditioning via rigidity losses is essential, not merely helpful.
CLSTM Improves Registration Accuracy Removing the stacked CLSTM and using a single-step registration approach reduced both tumor preservation and normal tissue alignment accuracy. The CLSTM's recurrent architecture enables capturing deformations at multiple scales and spatial locations progressively, with different encoder layers showing distinct activation patterns at different steps - coarser alignment early and finer refinement later.
Forward Direction Most Critical Ablation of individual loss components showed that tumor conditioning in the forward direction (preserving the moving image tumor) was more important than in the inverse direction for overall tumor preservation. However, removing either direction degraded performance, confirming the value of the full bidirectional approach.
Number of CLSTM Steps Matters Analysis showed that registration accuracy improved with increasing numbers of stacked CLSTM steps up to 8, with feature activations becoming more focused on regions of large anatomical difference (tumors, heart borders) in later steps. This confirms that the progressive refinement approach is more powerful than a single large deformation field computation.
Enabling Population-Level Dose-Outcome Studies TRACER enables voxel-based analysis studies that identify specific anatomical locations where high radiation dose correlates with complications or poor tumor control across patient populations. These studies have previously been limited by poor inter-patient registration quality, particularly in patients with large or centrally located tumors.
Automatic Pipeline Without Manual Segmentation Unlike previous tumor-aware registration methods that required manual tumor delineations as input, TRACER works with automatically segmented tumors from deep learning segmentation models. This makes it practically deployable in large-scale retrospective studies without requiring radiologist time for manual contouring of every scan.
Cross-Gender and Cross-Anatomy Robustness TRACER showed similar registration quality when aligning female to male patients or patients with tumors in different lobes - demonstrating robustness to the demographic and anatomical diversity present in real clinical populations. This consistency is critical for population studies where all patient-pair combinations must be registered.
Memorial Sloan Kettering Implementation The institutional Dataset III from Memorial Sloan Kettering Cancer Center demonstrates real-world clinical applicability. With NCI funding support, this work positions TRACER as a practical tool for the major cancer centers conducting large radiotherapy outcome studies, potentially advancing understanding of radiotherapy-related normal tissue toxicities.
Normal Tissue Accuracy Trade-off TRACER's tumor preservation comes at the cost of slightly lower normal tissue alignment accuracy compared to PACS. For applications where precise tubular organ (aorta, pulmonary artery) alignment is more important than tumor preservation, PACS might still be preferred. Future work should explore methods that achieve PACS-level normal tissue accuracy while maintaining TRACER's tumor preservation.
Downstream VBA Study Not Performed While TRACER demonstrates superior registration for VBA inputs, the study did not directly evaluate the quality of downstream VBA dose-outcome correlation findings. Whether TRACER's superior registration translates to better statistical power or new biological discoveries in VBA studies remains to be shown in dedicated VBA outcome studies.
Training Data Specificity TRACER was trained on locally advanced NSCLC patients treated with IMRT. Its performance for early-stage NSCLC with small or absent tumors, or other lung cancer treatment contexts (SBRT, proton therapy), has not been evaluated. Extending training to broader cancer contexts would expand applicability.
Future VBA Population Studies The authors explicitly identify comprehensive VBA population studies using TRACER as the primary next step. Such studies could identify specific lung, heart, or esophageal sub-regions where high radiation dose correlates with complications or recurrence, potentially leading to new treatment planning constraints that improve patient outcomes.