A Tensor-based Catheter and Wire Detection and Tracking Framework and Its Clinical Applications

First-in-human, real-time artificial intelligence assisted cerebral aneurysm coiling: a preliminary experience

BACKGROUND

Neuroendovascular procedures require careful and simultaneous attention to multiple devices on multiple screens. Overlooking unintended device movements can result in complications. Advancements in artificial intelligence (AI) have enabled real-time notifications of device movements during procedures. We report our preliminary experience with real-time AI-assisted cerebral aneurysm coiling in humans.

METHODS

A real-time AI-assistance software (Neuro-Vascular Assist, iMed technologies, Tokyo, Japan) was used during coil embolization procedures in nine patients with an unruptured aneurysm. The AI system provided real-time notifications for 'coil marker approaching', 'guidewire movement', and 'device entry' on biplane fluoroscopic images. The efficacy, accuracy, and safety of the notifications were evaluated using video recordings.

RESULTS

The AI system functioned properly in all cases. The mean number of notifications for coil marker approaching, guidewire movement, and device entry per procedure was 20.0, 3.0, and 18.3, respectively. The overall precision and recall were 92.7% and 97.2%, respectively. Five of 26 true positive guidewire notifications (19%) resulted in adjustment of the guidewire back toward its original position, indicating the potential effectiveness of the AI system. No adverse events occurred.

CONCLUSIONS

The software was sufficiently accurate and safe in this preliminary study, suggesting its potential usefulness. To the best of our knowledge, this is the first reported use of a real-time AI system for assisting cerebral aneurysm coiling in humans. Large scale studies are warranted to validate its effectiveness. Real-time AI assistance has significant potential for future neuroendovascular therapy.

Real time artificial intelligence assisted carotid artery stenting: a preliminary experience

BACKGROUND

Neurointerventionalists must pay close attention to multiple devices on multiple screens simultaneously, which can lead to oversights and complications. Artificial intelligence (AI) has potential application in recognizing and monitoring these devices on fluoroscopic imaging.

METHODS

We report out preliminary experience with a real time AI assistance software, Neuro-Vascular Assist (iMed technologies, Tokyo, Japan), in six patients who underwent carotid artery stenting. This software provides real time assistance during endovascular procedures by tracking wires, guiding catheters, and embolic protection devices. The software provides notification when devices move out of a predefined region of interest or off the screen during the procedure. Efficacy, safety, and accuracy of the software were evaluated.

RESULTS

The software functioned well without problems and was easily used. Mean number of notifications per procedure was 21.0. The mean numbers of true positives, false positives, and false negatives per procedure were 17.2, 3.8, and 1.2, respectively. Precision and recall were 82% and 94%, respectively. Among the 103 true positive notifications, 24 caused the operator to adjust the inappropriate position of the device (23%), which is approximately four times per procedure. False notifications occurred because of false positive device detection. No adverse events related to the software occurred. No periprocedural complications occurred.

CONCLUSIONS

Neuro-Vascular Assist, a real time AI assistance software, worked appropriately and may be beneficial in carotid artery stenting procedures. Future large scale studies are warranted to confirm.

Enhancing transvenous lead extraction risk prediction: Integrating imaging biomarkers into machine learning models

BACKGROUND

Machine learning (ML) models have been proposed to predict risk related to transvenous lead extraction (TLE).

OBJECTIVE

The purpose of this study was to test whether integrating imaging data into an existing ML model increases its ability to predict major adverse events (MAEs; procedure-related major complications and procedure-related deaths) and lengthy procedures (≥100 minutes).

METHODS

We hypothesized certain features-(1) lead angulation, (2) coil percentage inside the superior vena cava (SVC), and (3) number of overlapping leads in the SVC-detected from a pre-TLE plain anteroposterior chest radiograph (CXR) would improve prediction of MAE and long procedural times. A deep-learning convolutional neural network was developed to automatically detect these CXR features.

RESULTS

A total of 1050 cases were included, with 24 MAEs (2.3%) . The neural network was able to detect (1) heart border with 100% accuracy; (2) coils with 98% accuracy; and (3) acute angle in the right ventricle and SVC with 91% and 70% accuracy, respectively. The following features significantly improved MAE prediction: (1) ≥50% coil within the SVC; (2) ≥2 overlapping leads in the SVC; and (3) acute lead angulation. Balanced accuracy (0.74-0.87), sensitivity (68%-83%), specificity (72%-91%), and area under the curve (AUC) (0.767-0.962) all improved with imaging biomarkers. Prediction of lengthy procedures also improved: balanced accuracy (0.76-0.86), sensitivity (75%-85%), specificity (63%-87%), and AUC (0.684-0.913).

CONCLUSION

Risk prediction tools integrating imaging biomarkers significantly increases the ability of ML models to predict risk of MAE and long procedural time related to TLE.

Endovascular Tool Segmentation with Multi-lateral Branched Network during Robot-assisted Catheterization

Robot-assisted catheterization is routinely carried out for intervention of cardiovascular diseases. Meanwhile, the success of endovascular tool navigation depends on visualization and tracking cues available in the robotic platform. Currently, real-time motion analytics are lacking, while poor illumination during fluoroscopy affects existing physics- and learning-based methods used for tool segmentation. A multi-lateral branched network (MLB-Net) is herein proposed for tool segmentation in cardiovascular angiograms. The model has an encoder with multi-lateral separable convolutions and a pyramid decoder. Model training and validation are done on 1320 angiograms obtained during robot-assisted catheterization in rabbit. Model performance, explained with F1-score of 89.01% and mean intersection-over-union of 90.05% on 330 frames, indicates the model's robustness for guidewire segmentation in angiograms. The MLB-Net offers better performance than the state-of-the-art segmentation models such as U-Net, U-Net++ and DeepLabV3. Thus, it could provide basis for endovascular tool tracking and surgical scene analytics during cardiovascular interventions.

Guidewire Endpoint Detection Based on Pixel Adjacent Relation in Robot-assisted Cardiovascular Interventions

Visualization of endovascular tools like guidewire and catheter is essential for procedural success of endovascular interventions. This requires tracking the tool pixels and motion during catheterization; however, detecting the endpoints of the endovascular tools is challenging due to their small size, thin appearance, and flexibility. As this still limit the performances of existing methods used for endovascular tool segmentation, predicting correct object location could provide ways forward. In this paper, we proposed a neighborhood-based method for detecting guidewire endpoints in X-ray angiograms. Typically, it consists of pixel-level segmentation and a post-segmentation step that is based on adjacency relationships of pixels in a given neighborhood. The latter includes skeletonization to predict endpoint pixels of guidewire. The method is evaluated with proprietary guidewire dataset obtained during in-vivo study in six rabbits, and it shows a high segmentation performance characterized with precision of 87.87% and recall of 90.53%, and low detection error with a mean pixel error of 2.26±0.14 pixels. We compared our method with four state-of-the-art detection methods and found it to exhibit the best detection performance. This neighborhood-based detection method can be generalized for other surgical tool detection and in related computer vision tasks.Clinical Relevance- The proposed method can be provided with better tool tracking and visualization systems during robot-assisted intravascular interventional surgery.