Abstract
This paper addresses the need for improved CT-guidance during needle-based liver procedures (i.e., tumor ablation), while reducing the need for contrast agent injection during such interventions. To achieve this objective, we augment the intraoperative CT with the preoperative vascular network deformed to match the current acquisition. First, a neural network learns local image features in a non-contrasted CT image by leveraging the known preoperative vessel tree geometry and topology extracted from a matching contrasted CT image. Then, the augmented CT is generated by fusing the labeled vascular tree and the non-contrasted intraoperative CT. Our method is trained and validated on porcine data, achieving an average dice score of 0.81 on the predicted vessel tree instead of 0.51 when a medical expert segments the non-contrasted CT. In addition, vascular labels can also be transferred to provide additional information.
Introduction
Needle-based interventions, such as tumor ablation, are critical for treating liver cancer. These procedures rely heavily on Computed Tomography (CT) guidance to navigate the organ's complex vascular network. Traditionally, clinicians must inject contrast agents to visualize blood vessels and tumors. However, these agents are transient and potentially nephrotoxic, limiting their use during lengthy procedures that require multiple "check-point" scans.
Without contrast, the liver appears as a homogeneous mass to the human eye, making it difficult for surgeons to account for organ deformation caused by respiration or needle pressure. While rigid image fusion (overlaying preoperative scans) is a common workaround, it fails to capture non-rigid deformations. This research proposes a deep-learning-based pipeline to augment intraoperative non-contrast CT (NCCT) scans with preoperative vascular data, providing a high-accuracy, real-time "augmented" view without additional contrast.
Methodology
The proposed approach integrates a patient-specific training phase into the clinical workflow. By training on the subject's own anatomy prior to the main intervention, the system can better predict individual vascular shifts during the procedure.
The Augmentation Pipeline
The core of the method is a U-Net architecture designed to detect "vessel fingerprints"—subtle image features in NCCT that correspond to vascular structures.
- Vessel Map (VM) Extraction: A "Vessel Map" is generated from preoperative contrasted scans to define the region of interest for the vascular tree.
- Data Augmentation: To handle limited clinical data, hundreds of liver deformations are simulated using random displacement fields, teaching the network to recognize warped vascular structures.
- Neural Network Prediction: The U-Net receives two inputs: the preoperative vessel map and the intraoperative NCCT. It predicts the deformed position of the vessels by identifying sub-visual features in the non-contrasted data.
- Refinement and Labeling: Topological consistency is ensured via the Elastix library. Finally, anatomical labels are transferred from preoperative data to the intraoperative view.
Results
The method was validated using abdominal porcine data from four subjects. The results demonstrate a significant improvement over manual human performance and standard registration techniques:
- Predictive Accuracy: The system achieved a mean Dice score of 0.81 for vessel prediction.
- Human Comparison: In contrast, medical experts achieved a Dice score of only 0.51 when attempting to segment vessels on the same non-contrasted images.
- Importance of Priors: An ablation study showed that removing the preoperative Vessel Map from the input caused performance to drop to a Dice score of 0.16, confirming that preoperative geometry is a vital "prior" for the network.
Discussion and Conclusion
The primary innovation of this work is the concept of the "vessel fingerprint." While vessels are nearly invisible to the human eye in non-contrast CTs, convolutional neural networks can detect subtle textures if guided by preoperative geometry. This patient-specific approach overcomes the limitations of general registration tools that struggle with the uniform appearance of liver parenchyma.
Future work will focus on clinical validation in human patients and expanding the dataset beyond the initial porcine subjects. Additionally, user studies with clinicians are necessary to quantify the impact of this augmentation on surgical speed and precision in live operating environments.
Source code is available at: https://github.com/Sidaty1/Intraoperative_CT_augmentation
Summary updated