TraceTrans: Translation and Spatial Tracing for Surgical Prediction

TL;DR

TraceTrans addresses the reliability and interpretability gap in medical image translation for post-operative prediction by jointly estimating a deformation field and a translated image. The two-stream encoder-decoder design enforces spatial correspondences between pre- and post-operative states without fixed references, using differentiable mutual information and a cross-domain constraint to align structures. Across face cosmetology and brain MRI datasets, TraceTrans outperforms baselines in structural fidelity and realism, with ablations confirming the value of the deformation stream and cross-domain constraint. The work advances interpretable surgical prediction by providing pixel-level spatial traces of anatomical changes, enabling more trustworthy clinical deployment.

Abstract

Image-to-image translation models have achieved notable success in converting images across visual domains and are increasingly used for medical tasks such as predicting post-operative outcomes and modeling disease progression. However, most existing methods primarily aim to match the target distribution and often neglect spatial correspondences between the source and translated images. This limitation can lead to structural inconsistencies and hallucinations, undermining the reliability and interpretability of the predictions. These challenges are accentuated in clinical applications by the stringent requirement for anatomical accuracy. In this work, we present TraceTrans, a novel deformable image translation model designed for post-operative prediction that generates images aligned with the target distribution while explicitly revealing spatial correspondences with the pre-operative input. The framework employs an encoder for feature extraction and dual decoders for predicting spatial deformations and synthesizing the translated image. The predicted deformation field imposes spatial constraints on the generated output, ensuring anatomical consistency with the source. Extensive experiments on medical cosmetology and brain MRI datasets demonstrate that TraceTrans delivers accurate and interpretable post-operative predictions, highlighting its potential for reliable clinical deployment.

Figures (12)

• Figure 1: TraceTrans employs an encoder with dual decoders to predict both spatial deformation and translated images. The deformation field provides pixel-level correspondences from the source to the target, ensuring structural alignment while adapting to the target distribution. No fixed reference image is required during training and inference.
• Figure 2: TraceTrans jointly predicts spatial deformations and post-operative translations, enabling pixel-level correspondences and traceable structural changes for interpretable surgical outcome prediction. The second row illustrates the translation process, where spatial shifts are composed from velocity fields, while the first row shows that the close alignment between edges of the deformed pre-operative based on the predicted deformation and predicted post-operative images confirms anatomical correspondence.
• Figure 3: Overview of the TraceTrans framework. An encoder extracts features from the pre-operative image, which are processed by two decoders: the pixel translation module generates the post-operative prediction, and the spatial transformation module predicts the deformation field. The deformation ensures structural correspondence between the pre-operative and post-operative domains, while discriminators and multi-level constraints jointly optimize translation fidelity and anatomical consistency.
• Figure 4: Qualitative comparison of our method and several baselines. The images in the first row and the second row are the results on face cosmetology dataset and brain MRI dataset, respectively. The first two columns show the input and ground truth images, the last column shows Sobel edge maps derived from the deformed source images using the predicted deformation field of our method.
• Figure 5: Illustration of correspondence comparison in the ablation study. The proposed two‑stream constrained network generates deformations that align with the translated images, whereas a similar single‑stream architecture fails to preserve this correspondence.