Method for performing a dedicated scan of a structure under investigation using an imaging technique

Abstract

The invention relates to a method for performing a dedicated scan (1) using an imaging method (3) of a structure to be examined, in particular the left atrial appendage of a heart, comprising the following method steps: S1) Pre-selecting an area to be examined, wherein the area to be examined includes the structure to be examined, S2) Taking an overview photograph of the area to be examined, S3) Transmitting the overview recording to an evaluation unit (4), S4) Determining the position of the structure to be investigated in the overview image using artificial intelligence and determining preliminary coordinates, wherein the preliminary coordinates define the position of the structure to be investigated in the overview image, S5) Determining final coordinates using artificial intelligence based on preliminary coordinates, wherein the final coordinates define the position of the structure to be examined for the imaging procedure, S6) Transmitting the final coordinates to a control unit of the imaging procedure, S7) Performing a dedicated scan of the structure to be investigated using the transmitted final coordinates via the imaging technique (3). In this way, a method for performing a dedicated scan is provided, which can also detect structures that are difficult to define and with which radiation exposure can be reduced.

Description

[0001] The invention relates to a method for performing a dedicated scan of a structure to be examined, in particular the left atrial appendage of a heart, using an imaging technique.

[0002] A dedicated scan refers to an imaging procedure, such as a computed tomography (CT) scan, that is specifically tailored to a particular area of ​​the body or a specific medical question. Unlike a general or comprehensive scan, which may cover multiple body regions, a dedicated scan focuses on a specific organ, structure, or medical issue. The goal of a dedicated scan is to produce detailed, high-resolution images of a specific area to enable an accurate diagnosis. By focusing on a particular area, image quality can be optimized by adjusting parameters such as resolution or contrast agent usage.Furthermore, targeted examination can minimize the radiation exposure for the patient, especially for radiation-sensitive organs at risk, as only the relevant area is scanned.

[0003] The left atrial appendage, also known as the left heart appendage or left atrial appendage (LAA), is a small, sac-like structure located at the left atrium of the heart. It plays a significant role in the development of certain heart conditions, particularly the formation of blood clots. One of the most common conditions affecting the left atrial appendage is atrial fibrillation, a heart rhythm disorder in which the atria beat irregularly and often very rapidly. This irregular contraction frequently causes blood to stagnate in the left atrial appendage, increasing the risk of blood clot formation. Such clots can travel to the systemic circulation and significantly increase the risk of stroke, making the left atrial appendage an important clinical focus. Another relevant problem associated with the left atrial appendage is thrombus formation.In cases of atrial fibrillation or other conditions that lead to reduced blood flow in the left atrial appendage, thrombi (blood clots) can form. These clots pose a significant risk because they can enter the bloodstream and trigger embolic events such as strokes. To minimize this risk, blood-thinning medications are often used. In certain cases where long-term anticoagulation is not possible, the left atrial appendage is also surgically closed or closed using a special device. This procedure, called left atrial appendage occlusion (LAA occlusion), prevents the formation and release of blood clots. Although less common, inflammatory diseases that can affect the left atrial appendage, such as endocarditis (inflammation of the inner lining of the heart), are also important. These inflammations usually affect the entire heart but can also involve the left atrial appendage.Congenital anomalies of the left atrial appendage are also rare, but can lead to abnormal function or an increased risk of thrombus formation and are often discovered in association with other congenital heart defects. Tumors of the left atrial appendage, such as primary cardiac tumors (e.g., myxomas) or metastases affecting the heart tissue, are extremely rare. Overall, the left atrial appendage is of clinical importance, particularly in the context of atrial fibrillation and the associated thrombus formation, as this poses a significant risk of stroke.

[0004] Traditionally, examinations specifically targeting the left atrial appendage involve a scan of the entire heart. This is because the left atrial appendage, due to its significant anatomical variability, is difficult to precisely delineate in the topogram used for planning procedures such as CT scans. Therefore, in clinical practice, medical personnel typically delineate the entire heart. However, the radiation exposure for patients is considerably higher with a whole-heart scan than with a dedicated scan of the left atrial appendage, as a larger area is scanned than actually necessary. Since any form of radiation exposure is considered potentially hazardous to health, this represents a significant disadvantage for patient safety.

[0005] Based on this, the object of the invention is to provide a method for performing a dedicated scan, which can also detect structures that are difficult to define and with which radiation exposure can be reduced.

[0006] This problem is solved by the subject matter of claim 1. Preferred embodiments are found in the dependent claims.

[0007] According to the invention, a method for performing a dedicated scan of a structure to be examined, in particular the left atrial appendage of a heart, using an imaging technique, is provided, comprising the following method steps: Pre-selecting an area to be examined, wherein the area to be examined includes the structure to be examined, Creating an overview image of the area to be examined, transmitting the overview image to an evaluation unit, Determining the position of the structure to be examined in the overview image using artificial intelligence and determining preliminary coordinates, wherein the preliminary coordinates define the position of the structure to be examined in the overview image, Determining final coordinates using artificial intelligence based on preliminary coordinates, where the final coordinates define the position of the structure to be examined for the imaging procedure, Transmitting the final coordinates to a control unit of the imaging process, Performing a dedicated scan of the structure to be examined using the transmitted final coordinates via the imaging procedure.

[0008] The imaging procedure preferably includes a three-dimensional scan, particularly preferably a CT scan, but does not exclude other imaging methods. The particular challenge lies not only in delineating the structure to be examined, but also in determining the correct coordinates for the CT scan. A direct transfer of coordinates from the overview image or topogram to the coordinates for the CT scan is not possible, as the dimensions of the images are different.

[0009] It is therefore a key aspect of the invention that artificial intelligence is used to delineate the structure to be examined and to determine suitable coordinates, wherein the artificial intelligence is trained to recognize the structure to be examined and to predict coordinates for the imaging procedure.

[0010] There are some structures in the body that cannot be identified in a standard overview image because a three-dimensional scan is required. This is the case, for example, with the left atrial appendage. The left atrial appendage can only be located using three-dimensional imaging. Therefore, the artificial intelligence is trained to locate the structure to be examined in the scan based on previously performed scans and corresponding overview images, and to transfer this position to the respective overview image. When the artificial intelligence then receives a new overview image, it can use the training data to predict where the structure to be examined, in particular the left atrial appendage, will be located.Furthermore, the artificial intelligence is trained not only to locate the structure to be examined based on the overview image, but also to determine suitable coordinates for the scan. This ensures that the structure under investigation is captured by the scan or imaging process without scanning additional volume outside the structure being examined. Even if the structure to be examined has been identified or predicted in the overview image, and its position in the overview image can be determined by coordinates, these coordinates cannot be used directly to determine the region that must be scanned by the imaging process. There are usually discrepancies in the coordinates between the overview image and the imaging process.Therefore, the artificial intelligence is also trained to detect deviations in coordinates between the overview image and the imaging procedure. A deviation could, for example, be a longitudinal shift of a few centimeters.

[0011] The artificial intelligence is trained primarily retrospectively. This means that annotations from the imaging procedure are preferably transferred back to the overview image or the topogram. This ensures that the topogram is annotated correctly and that patient movements, such as a beating heart, can be at least partially taken into account.

[0012] It has been shown that by using artificial intelligence for precise delineation, a reduction in radiation exposure of approximately 50% can be achieved compared to examining the entire area under investigation. For example, it has been shown that the radiation exposure during a dedicated scan of the left atrial appendage using the inventive method could be reduced by at least 5 mSv, or 55%, compared to a CT scan of the entire heart.

[0013] According to a preferred embodiment of the invention, the overview image comprises a two-dimensional tomographic image, in particular a topogram. A two-dimensional tomographic image or an X-ray image enables the rough delimitation of the area to be examined with relatively low radiation exposure.

[0014] According to a preferred embodiment of the invention, the final coordinates define the position of the structure to be examined, plus a safety margin for the imaging procedure. Due to patient movement, a safety margin is placed around the delimited volume, at least in the cranial-caudal direction, i.e., longitudinally. The spatial extent and size of the safety margin are determined by the type of anatomical structure being examined. For example, in the case of an anatomical structure of the heart, a safety margin of approximately 1 cm from the carina to the apex has proven useful.

[0015] According to a preferred embodiment of the invention, the artificial intelligence comprises deep learning. Deep learning is a special form of machine learning that uses multi-layered neural networks (so-called deep neural networks). It is particularly effective in processing large amounts of data, especially in image and speech recognition.

[0016] According to a preferred embodiment of the invention, the deep learning comprises a model type selectable from the following list: artificial neural network, convolutional neural network, region-based convolutional neural network, cascaded region-based convolutional neural network, task-aligned one-stage object recognition and / or verifocal network.

[0017] Artificial Neural Networks (ANNs) are used for general tasks such as classification and regression. These networks consist of an input layer, one or more hidden layers, and an output layer, with each layer consisting of neurons that operate with weights and activation functions.

[0018] Convolutional Neural Networks (CNNs) are particularly well-suited for processing image data. CNNs contain specialized layers such as convolution and pooling layers that extract features from images and are used for image classification, object detection, and image segmentation.

[0019] A region-based convolutional neural network (R-CNN) works in several steps: First, potential areas in the image that could contain an object (so-called region proposals) are generated. Then, features are extracted for each of these regions using a convolutional neural network (CNN). These features are then used to classify the object and determine its precise position in the image.

[0020] A Cascade R-CNN is an advanced deep learning model used in image processing, particularly object recognition. It is based on the concept of the Region-based Convolutional Neural Network (R-CNN) and extends it with a multi-stage architecture. The Cascade R-CNN uses a series of staged detectors that build upon each other to progressively improve the model's performance. The basic idea is to refine object recognition in multiple stages. The model starts with a rough prediction at the first stage, which is then further refined in the next stage. This process is repeated across several stages, with each subsequent stage aiming to correct the errors of the previous stage. This continuously improves the accuracy of the recognition.A key advantage of the Cascade R-CNN is its ability to reduce the problem of overfitting, which often occurs when a single detector is trained for very precise detection. Cascading allows the model to address specific weaknesses at each stage, thereby achieving higher overall accuracy.

[0021] The Cascade R-CNN is used in many application areas, such as surveillance, autonomous vehicles, medical imaging, and anywhere precise object recognition in images is required. Additionally, it can be used for tasks that require precise localization or segmentation of objects within an image. Overall, the Cascade R-CNN represents a significant advancement over traditional R-CNN approaches, as it progressively corrects the weaknesses of earlier predictions, thus achieving higher precision and robustness in object recognition.

[0022] TOOD, which stands for "Task-aligned One-stage Object Detection," is an object detection model designed to improve the performance and efficiency of one-stage detectors. TOOD's primary goal is to overcome the challenges of conventional one-stage detectors, particularly regarding the alignment of classification and localization tasks. In object detection, models typically consist of two main components: classification, where the model identifies the class to which an object belongs, and localization, where the object's precise position within the image is determined. Conventional one-stage detectors, such as YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector), often struggle to perform these two tasks accurately and efficiently because they are handled independently.This often leads to suboptimal performance because the classification and localization results are not optimally aligned. The TOOD model was developed to solve this problem by better aligning the classification and localization tasks in a single-stage detector. This is achieved through a joint learning process that ensures the results of these two tasks are more closely matched and mutually reinforcing. A key feature of TOOD is task alignment, where the classification and localization tasks are adjusted to optimize each other. This results in improved overall performance, as both tasks benefit from each other. Another important feature of TOOD is the use of an adaptive loss function.This feature dynamically balances classification and localization errors, depending on which aspect requires more improvement. This results in finer tuning and better overall model performance. Because TOOD is a one-step detector, it scans the image directly in a single pass and detects objects without requiring a separate region suggestion phase. This makes TOOD faster and more efficient than many two-step detectors that use a region suggestion phase. Thanks to the improved balance between classification and localization tasks, TOOD achieves high accuracy and shows a remarkable performance increase compared to traditional one-step detectors.

[0023] VFNet, which stands for "Varifocal Net," is an object detection model specifically designed to improve the accuracy and reliability of object detection and localization in images. VFNet belongs to the class of one-stage detectors, which aim to detect objects in a single pass across the image, making it faster and more efficient than two-stage detectors. A key feature of VFNet is the introduction of varifocal loss, a special loss function designed to optimize the prediction of classification probabilities and the quality of localization. In conventional object detection models, there is often a discrepancy between the classification probability and the actual accuracy of the localization.VFNet attempts to solve this problem by directly linking the classification probability to the quality of the predicted bounding box. The varifocal loss ensures that the model gives greater weight to both classification accuracy and localization precision when calculating the loss function. This leads to a better alignment of these two aspects and thus to overall more precise object detection. VFNet also employs a centerness-aware quality estimation branch, which helps reduce uncertainty in the predictions and ensures more stable and accurate object detection. This method improves the model's ability to focus on the relevant regions in the image and contributes to increased localization precision.Another advantage of VFNet is its seamless integration into existing object detection architectures, requiring no significant network architecture changes. This makes it a flexible and versatile solution for various applications where fast and accurate object detection is essential.

[0024] According to a preferred embodiment of the invention, the artificial intelligence was trained with the following training data: bounding boxes, labeled images, image pairs and / or pixel-accurate annotations.

[0025] Labeled images are defined as image files with corresponding labels (e.g., "dog," "cat") for classification tasks. Image pairs are used particularly for image-to-image translation. Pixel-perfect annotations are used particularly for segmentation tasks where each pixel must be assigned to a specific object or class.

[0026] The Cascade R-CNN, TOOD, and VFNet models are advanced object recognition models that require specific training data to perform their tasks effectively. All three models rely on image data depicting various objects in different scenes and contexts. This imagery must be sufficiently diverse to ensure the models are able to recognize objects in different environments and under varying conditions. A crucial component of the training data is precisely annotated bounding boxes that specify the exact positions of the objects within the image. These bounding boxes are essential because the models need to learn how to accurately locate and delineate objects within an image.

[0027] In addition to the bounding boxes, the models require labels for each object within these boxes. These labels indicate the type of object, such as "vessel," "heart chamber," or "atrium." These classification labels are essential for the models' recognition tasks. For specialized applications, such as Cascade R-CNN, additional annotations like segmentation masks may be necessary to segment objects at the pixel level.

[0028] The TOOD model also requires a large number of images with annotated bounding boxes and labels. A unique feature of TOOD is its ability to benefit from quality ratings that assess the quality of the bounding boxes. These ratings help the model distinguish between high-quality and lower-quality suggestions, which is particularly relevant when applying TOOD's adaptive loss function.

[0029] The VFNet model also requires image data with annotated bounding boxes and labels. A special feature of VFNet is its use of uncertainty information that may be contained in the training data. This information helps the model optimize prediction quality by better aligning classification probabilities with localization accuracy.

[0030] Overall, a large amount of diverse and precisely annotated image data is provided, particularly for the three aforementioned models, to fully exploit the potential of these complex models. Data quality plays a crucial role, as inaccurate or erroneous data can lead to poorer model performance. A wide range of scenarios, lighting conditions, perspectives, and backgrounds in the training data ensures that the models are robust against variations in the input data, thus achieving high accuracy and reliability in object recognition.

[0031] It has been shown that, particularly with the three models Cascade R-CNN, TOOD and / or VFNet, a high accuracy in the prediction of the structure to be investigated, preferably approximately 98%, can be achieved.

[0032] According to the invention, the use of a method described above for the dedicated scanning of a left atrial appendage of a heart using an imaging method is further provided.

[0033] The invention will now be explained in more detail with reference to the drawings and a preferred embodiment.

[0034] The drawings show Fig. 1 schematically a flowchart of a process according to a preferred embodiment of the invention.

[0035] The procedure will be in Fig.1. The imaging procedure 3 is described using the example of the left atrial appendage of a heart to be examined in a CT scan. The left atrial appendage is particularly difficult to delineate in a two-dimensional overview image. In order to nevertheless perform a dedicated CT scan, the patient is positioned in the gantry. Subsequently, the radiology staff determines a region to be examined (S1) and specifies this region to the CT control software. An overview image is then acquired (S2). This overview image is transmitted to the evaluation unit (S3). After the overview image has been received, artificial intelligence determines the position of the structure to be examined, in this case the left atrial appendage (S4), and the coordinates of the structure to be examined in the overview image are converted into coordinates for the CT scan (S5).These final CT coordinates are then transmitted to the CT scanner's control software (S6). Subsequently, the dedicated CT scan is performed using these final coordinates (S7). The process is divided into three areas: steps for the radiology staff (2), steps for the imaging procedure (3) or the control software, and steps for the evaluation unit (4) with its artificial intelligence. The evaluation unit (4), or the artificial intelligence, identifies the structure to be examined and determines suitable CT coordinates so that the dedicated CT scan can be performed using these coordinates. By reducing the volume to be examined to the structure itself, possibly with an additional safety margin, the radiation dose can be halved. Reference symbol list 1. Procedure for performing a dedicated scan by radiology professionals 2 3 Imaging procedure 4 evaluation units S1 Preselecting an area to be examined S2 Taking an overview photograph of the area to be examined S3 Transmitting the overview image to an evaluation unit S4 Determining the position of the structure to be examined in the overview image using artificial intelligence and determining preliminary coordinates S5 Determining final coordinates using artificial intelligence based on preliminary coordinates S6 Transmitting the final coordinates to a control unit of the imaging procedure S7 Performing a dedicated scan of the structure under investigation

Claims

[1] Method for performing a dedicated scan (1) using an imaging technique (3) of a structure to be examined, in particular the left atrial appendage of a heart, comprising the following procedural steps: S1) Pre-selecting an area to be examined, wherein the area to be examined includes the structure to be examined, S2) Taking an overview photograph of the area to be examined, S3) Transmitting the overview recording to an evaluation unit (4), S4) Determining the position of the structure to be investigated in the overview image using artificial intelligence and determining preliminary coordinates, wherein the preliminary coordinates define the position of the structure to be investigated in the overview image, S5) Determining final coordinates using artificial intelligence based on preliminary coordinates, wherein the final coordinates define the position of the structure to be examined for the imaging procedure, S6) Transmitting the final coordinates to a control unit of the imaging procedure, S7) Performing a dedicated scan of the structure to be investigated using the transmitted final coordinates via the imaging technique (3). [2] Method according to claim 1, wherein the overview image comprises a two-dimensional layer image, in particular a topogram. [3] Method according to claim 1 or 2, wherein the final coordinates define the position of the structure to be examined plus a safety margin for the imaging method (3), [4] Method according to any of the preceding claims, wherein the artificial intelligence comprises deep learning. [5] Method according to any of the preceding claims, wherein the deep learning comprises a model type selectable from the following list: artificial neural network, convolutional neural network, recurrent neural network, region-based convolutional neural network, cascaded region-based convolutional neural network, task-aligned one-stage object recognition and / or verifocal network. [6] Method according to any of the preceding claims, wherein the artificial intelligence was trained with the following training data: bounding boxes, labeled images, image pairs and / or pixel-accurate annotations. [7] Use of a method (1) according to any of the preceding claims for dedicated scanning of a left atrial appendage of a heart using an imaging method.

Images