Trauma ultrasound tissue-mimicking phantom

Abstract

A phantom (100) used in a) accumulating ultrasound images for the purpose of training Al models and b) training medical personnel to conduct ultrasound scans for eFAST examinations. The phantom (100) includes a body (101) in the shape of a human torso having a plurality of cavities formed in it for organs and fluid bladders to simulate injuries. The phantom (100) further including an actuator (532, 1332) for causing movement of a simulated lung (or lung segment) within a rib cage area to simulate breathing.

Description

Trauma Ultrasound Tissue-Mimicking PhantomI. Field of the Invention

[0001] The invention relates to a tissue-mimicking phantom that can be used in training medical personnel including first responders and artificial intelligence (Al) models to perform an extended-focused assessment with sonography for trauma (eFAST) examination. In at least one embodiment, the phantom includes ballistic gel formed as a phantom in a torso mold with cavities created for a rib cage, simulated lungs, and different organs. In at least one embodiment, the phantom includes tillable bags to simulate injuries within the torso area around organs.II. Background and Overview Summary

[0002] Ultrasound (US) imaging is widely used in combat casualty care and emergency medicine for diagnosing and triaging trauma patients and prioritizing casualty evacuations. One of its most common uses is the eFAST exam, which focuses on scanning specific anatomical regions to identify free fluid or air as an indicator of injury. This use of US enables the detection of abdominal hemorrhage (AH), pneumothorax (PTX), and hemothorax (HTX) injuries.

[0003] While eFAST exams are essential for emergency diagnosis, they can require extensive training to appropriately position the US transducer and interpret the US images. Thus, skilled medical professionals are required to successfully conduct an eFAST exam and accurately interpret the US images. This necessary medical training is complex and has had limited implementation in standard military training. To accomplish the goal of making US imaging viable in battlefield settings, the integration of Al can enhance eFAST examination specifically through providing (i) guidance to properly image scan sites through anatomical recognition and (ii) automating image interpretation and diagnoses from captured images.

[0004] In order to develop deep learning Al models for medical imaging applications, large data sets are required that contain images of the relevant anatomy both with and without injury. Obtaining this type of training data can be challenging in humans due to the time sensitive nature of these traumatic injuries. Similarly, capturing this data in an animal model yields other time and cost challenges that fail to serve as a practical substitute to human studies. Thus, tissuemimicking phantoms function as an alternative. There are multiple commercially available FAST and eFAST trainers with varying adjustable injury states and image quality capabilities. For example, the SonoSkin Trainer from Simulab Corporation utilizes a life-like cover that can be scanned with a mock US probe and pre-programmed images are shown. This trainer is very limited for Al development as there is no variance in the images, and the software does not account for varying probe angles during an eFAST exam. Other US compliant phantoms have been developed to overcome this limitation, such as the Blue Phantom FAST trainer from CAE Healthcare that is made of US compliant materials and can simulate hemorrhage by inserting fluid into the heart, spleen, bladder and liver scan sites. However, since the fluid cannot be removed completely, negative injury scans are not able to be collected for model training.Additionally, the phantom lacks a thoracic cavity with anatomically relevant features making diagnostic data collection for thoracic-related injuries impossible.

[0005] Tissue phantoms, if properly developed, can accelerate automation technology development by allowing initial Al model training and troubleshooting to be possible without the need for animal or human testing beyond what is currently possible.III. Summary of the Invention

[0006] With this in mind, a torso tissue phantom has been created that can mimic PTX and HTX (left and right) injuries and AH injuries (e.g., left upper quadrant (LUQ), right upper quadrant (RUQ), and pelvis (BLD) injuries) and allow for data acquisition for development of Al models and to train medical personnel. An US compatible thoracic tissue phantom simulating lung motion, with modular HTX and PTX injuries, supporting the development of Al image interpretation models was developed. In at least one embodiment, the phantom may be used for training personnel on how to detect HTX, PTX, and / or AH injuries.

[0007] In at least one embodiment, the phantom successfully integrated injury sites for four out of the five scan areas evaluated during the exam. The PTX methodology built on a simple 2D phantom that was previously developed by integrating a similar motion concept to mimic lung motion in a realistic rib cage. Resulting images from that embodiment resemble those from the prior phantom and human US scans. However, the methodology in its current form lacks the capability of altering respiratory rate in baseline images and the creation of lung points which are often looked for when diagnosing PTX clinically. HTX was accurately detected at the right and left side scan points and mimicked human US images, while AH was detectable across three abdominal scan points (LUQ, RUQ, BLD). HTX and AH can be visualized simultaneously or independently at a scan point to create more variations in the phantom setup when simulating an eFAST exam. The bladder volume present at the pelvic scan point was modular to represent a more and less full bladder as the effect of that on successfully identifying AH at this scan point is well known (Richards and McGahan, 2017; Rowland-Fisher and Reardon, 2021). The only eFAST scan point not included was the subxiphoid view for cardiac assessment. Creating an US phantom analogue that mimicked heart motion and had realistic heart chamber structure for proper eFAST examination was not possible with the setups needed at the RUQ and LUQ viewpoints and PTX scan sites. Tissue phantoms exist for the cardiac view and even hemopericardium detection which could be used in conjunction with this tissue phantom to allow for inclusion of this eFAST scan point (for example, phantoms from CAE Healthcare, 2022).

[0008] To demonstrate a use case for the phantom, it was evaluated with existing or newly trained Al models for the various scan points using images collected in the eFAST protocol. The new models for AH and HTX for three different scan points were successful at accurately detecting these injuries at more than 95% accuracy for all instances. While one of the benefits for the tissue phantom is that these different viewpoints can be reformatted to increase subjectvariability, it is still less than biological noise, so these high performances were expected. However, efforts were taken during Al model training to prevent overfitting. A widely used approach is to augment image inputs so that Al models less easily focus on image artifacts not associated with the injury. For this effort, we used rotation, flip, zoom, translation augmentations similar to approaches successfully used in other Al US imaging efforts (Hussain, (2017), “Differential data augmentation techniques for medical imaging classification tasks,” AMIA annual symposium proceedings, Am. Medical Informatics Assoc., 979.; Xu, (2022), “A comprehensive survey of image augmentation techniques for deep learning,” preprint; Snider, (2023), “Using ultrasound image augmentation and ensemble predictions to prevent machinelearning model overfitting,” Diagnostics 13, 417). Another approach taken was to include validation patience during training so that training ceased if validation loss did not decrease for five training epochs. This prevents model overfitting during 100s of training epochs and was triggered within 20-40 epochs for all training performed. With the limited datasets at this point, the ultimate validation was the Gradient Class Activation Maps (GradCAM) overlays that highlight the region on the image that is driving predictions. With the described overfitting prevention methods, GradCAM overlays showed the majority of predictions were tracking proper injury locations. However, some images continue to track image artifacts not associated with injury, so more training data and subject variability will be needed to further improve these developed Al models. Regardless, this use case still highlights the potential use for this phantom.

[0009] Unlike the other injury conditions, a model for PTX detection from segments of M-Mode images which was successful for blind PTX detection in swine images was previously developed (Boice, (2022), “Training ultrasound image classification deep-learning algorithms for pneumothorax detection using a synthetic tissue phantom apparatus," J. Imaging 8, 249). This deep-learning model was only trained on a simple PTX tissue phantom. Using this trained model, prediction performance was heterogeneous based on US imaging system. Specifically, images from one system were 50% accurate, with every prediction failing to see a PTX injury, while the other US systems were more than 85% accurate, achieving slightly lower performance than the swine image predictions at 93% (Boice). Image acquisition bias on Al model predictions is a known issue that includes medical imaging equipment bias among other potential biases that can limit model generalization (Drukker, (2023), “Toward fairness in artificial intelligence for medical image analysis: identification and mitigation of potential biases in the roadmap from data collection to model deployment,” JMI 10, 061 104). The availability of US technology differs, and it is always advancing, so it is imperative that instrument noise be accounted for during training to make Al models for eFAST be more robust for implementation on multiple platforms.

[0010] In at least one embodiment, a method of making a phantom including pouring ballistics gel into a cast mold made from a silhouette, once an initial layer of ballistic gel was poured, placing a resin printed rib cage with a solid cavity into the mold and the remaining ballistic gel ispoured in to completely fill the cast mold, removing the inner cavity, carving an access hole in the head / neck region, optionally inserting a plurality of organs into the phantom body after carving out cavities for the organs, optionally connecting a lung phantom having a sled into the rib cage, and optionally attaching the sled to a reciprocating linear actuator. The method further including placing at least one fluid or air bladder configured to be fillable between the lung and a thoracic area of the phantom to create modular injuries and / or at least one bladder proximate one organ, and optionally causing movement of the lung phantom by activation of the reciprocating linear actuator where optionally the movement of the lung is along a longitudinal axis. The method according to any of the above embodiments where the lung is made of foam and / or for each organ, pouring ballistic gel into a mold for each organ. The method according to any of the above embodiments, further including printing the sled out of resin

[0011] In at least one embodiment, a phantom including: a body made of ballistics gel, optionally the body is in the form of a human torso; a rib cage made of resin inside of the body; a sled within a cavity of the rib cage, the sled optionally having a base on which two opposed walls extend away with one wall having a connection mount; a lung made of foam inside of the rib cage and attached to the sled, optionally the lung sits on the optional base and between the optional walls of the sled, optionally the foam is aerated to mimic breathing motion during operation; and an expiration system connected to the sled and the expiration system configured to cause movement of the sled where adjustment of the frequency leads to an adjustment in breath rate with a slower frequency corresponding to a decreased breath rate, optionally the expiration system connects to the optional connection mount. In a further embodiment, the expiration system includes a) a reciprocating linear actuator connected to the lung or b) a support structure having a vertical wall, a reciprocating linear actuator mounted on the support structure and connected to the lung, a control arm mounted on an actuator shaft of the actuator to rotate about the actuator shaft, a linkage arm connected to the control arm, and a shaft connected to the linkage arm and the sled and configured to provide longitudinal movement of the sled in the rib cage. Further to the above phantom embodiments, the phantom further including at least one fluid fill bladder placed between the lung and the body with variable volume to replicate different injury statuses, and / or at least one air fill bladder placed between the lung and a thoracic area of the body with variable volume to replicate different injury statuses.

[0012] Further to the above phantom embodiments, the phantom further including a plurality of organs, which optionally are made of ballistic gel and / or naturally dried gelatin. Further to the other embodiments in this paragraph, a) the organs are selected from a diaphragm, a spleen, a liver, a stomach, a kidney, a bladder, a rectum, and an intestine or b) the organs include a spleen, a liver, a kidney, and a bladder. Further to the other embodiments in this paragraph, at least one organ bladder proximate to at least one organ. In a further embodiment, the organ bladder is fillable with air or fluid to press up against the at least one organ that it is near.

[0013] To overcome some of the challenges posed by limited trained personnel and the difficulties of US imaging in battlefield settings, the phantom according to at least one embodiment was developed. The tissue phantom was designed to simulate multiple thoracic and abdominal injuries that are routinely assessed during eFAST exams. It incorporates realistic anatomy and simulation of injury to support Al anatomical guidance and injury diagnostic models to determine injury status and severity.IV. Brief Description of the Drawings

[0014] FIG. 1A illustrates a mold for the phantom according to at least one embodiment of the invention. FIG. 1 B illustrates the phantom with hollowed out cavities in the mold according to at least one embodiment of the invention.

[0015] FIG. 2 illustrates a method for making a phantom according to at least one embodiment of the invention.

[0016] FIG. 3 illustrates examples of the organs and some of their molds for use in the phantom according to at least one embodiment of the invention.

[0017] FIG. 4 illustrates an example of the sled according to at least one embodiment of the invention.

[0018] FIGs. 5A and 5B illustrate a CAD of the setup for the lung region of the phantom according to at least one embodiment of the invention. FIG. 5C illustrates a built phantom with the scan areas identified according to at least one embodiment of the invention.

[0019] FIG. 6 illustrates a Stress-Strain plot of the respective 100% Clear Ballistic Gel (CBG), 70%, 50% CBG. The Stress-Strain plots of the tested CBG mixtures are shown above, 100% CBG performed best under testing.

[0020] FIGs. 7A and 7B illustrate example bladders according to at least one embodiment of the invention.

[0021] FIG. 8 illustrates representative M-mode US images for Negative, PTX, and HTX thoracic conditions at different breathing rates achieved by the expiration system.

[0022] FIG. 9 illustrates representative B-mode US images for the HTX and LUQ scan points showing increasing injury volume. Images go from no injury to a full injury from left to right.

[0023] FIG. 10 illustrates the phantom with a rib cage and pink lung attached to the actuator according to at least one embodiment of the invention.

[0024] FIG. 1 1 illustrates a close-up of the shoulder area of the phantom illustrated in FIG. 10.

[0025] FIG. 12 illustrates representative US images for each injury condition at each scan point with GradCAM overlays.

[0026] FIG. 13 illustrates a 3D rendering of all body parts used in a second phantom embodiment.

[0027] FIG. 14 illustrates a diagram of the mechanism used to simulate a breathing lung in US for the second phantom embodiment.

[0028] FIG. 15 illustrates example US images after image augmentation has been randomly done for (A) pelvic view, (B) right upper quadrant view, and (C) left upper quadrant view images.

[0029] FIGs. 16A-16D illustrate US scans of a recreation of a pneumothorax injury in the second phantom embodiment.

[0030] FIGs. 17A-17D illustrate Confusion Matrices and US scans for PTX model predictions on test M-mode images collected in the second phantom embodiment.

[0031] FIG. 18 illustrates GradCAM overlays for PTX Al model predictions on test images collected using the second phantom embodiment.

[0032] FIGs. 19A-19E illustrate a pelvic view in the phantom and associated US images.

[0033] FIGs. 20A-20B illustrate performance results for an Al classification model trained for the pelvic eFAST view and US scans.

[0034] FIGs. 21-21 H illustrate 3D representation of the phantom and US scans for a RUQ view in the phantom or a human.

[0035] FIGs. 22A-22D illustrate performance results for an Al classification model trained for the RUQ eFAST view.

[0036] FIGs. 23A-23H illustrate 3D representation of the phantom and US scans for a LUQ view in the second phantom embodiment.

[0037] FIGs. 24A-24D illustrate performance results for an Al classification model trained for the LUQ eFAST view.V. Detailed Description

[0038] The tissue phantom was developed as a necessity for ongoing research efforts as no existing commercial phantom on the market was capable of creating US compliant positive and negative injury states for both abdominal and thoracic injuries - all of which are critical for eFAST triage procedures.

[0039] FIGs. 1A-1 B, 13, and 14 illustrate two phantoms 100, 1300 with different components and configurations. FIGs. 1A-1 B illustrate a mold 190 to form the phantom body 101. The mold 190 was created by casting a plastic mannequin silhouette. The mold 190 is configured to go onto a table 192 that includes rotating components 194 on the end as illustrated in FIG. 1 B that includes a phantom body 101. FIG. 2 illustrates a manufacturing method for the phantom 100.

[0040] The phantom 100 was created by pouring clear 10% ballistic gel (CBG10; Clear Ballistics, Greenville, SC, USA) containing 1% w / w silica gel (Sigma-Aldrich, St. Louis, MO, USA) into the mold 190, 205. When pouring the CBG10, a 3D printed ribcage, printed with High Temperature Resin (FormLabs, Somerville, MA, USA), with an embedded 3D printed polylactic acid mold was placed inside, 210. In at least one embodiment, the rib cage with a thorax placeholder is placed while the layer is still malleable. Once the phantom body 101 was fully poured and cooled, 215, the back of the phantom body 101 was carefully carved and the 3D printed mold was taken out, leaving a hollow rib cavity 102 as illustrated in FIG. 1 B, 220. In atleast one embodiment, an access hole 103 was carved from the head / neck region for the control rod to passthrough. Three more cavities were made for the right upper quadrant (RUQ) 104, left upper quadrant (LUQ) 105, and pelvis (BLD) 106 at their corresponding anatomical sites, 225. Internal organs were created by pouring the same ballistic gel and silica gel mixture into molds for kidneys, liver, and spleen. Bladder and rectum were poured using CBG 10 without silica gel, 230. FIG. 3 illustrates different organs and sample molds. The simulated lung tissue was created by pouring Soma Foama™ 25 (Smooth-On Inc, Macungie, PA, USA) into the hollow ribcage cavity 102, 235. Once cured, the foam lung 520 was removed and shaved to allow the lung to slide smoothly inside the cavity against the cavity walls, 240. The foam lung 520 was then attached to a sled 410 (FIG. 4), which may be 3D printed or casted, along a T-rail of the expiration system 540 using a 160 reps per minute reciprocating linear actuator 532 (Amazon, Seattle, WA, USA), 245, as illustrated in FIGs. 1 B and 5A-5B.

[0041] CBG 10 was chosen as the main material for casting the phantom due to its durability, since the thoracic phantom must withstand continuous forces exerted during an US examination and lung motion. The expiration via the reciprocating linear actuator 532 applies both tensile and compressive forces as it moves through the thoracic cavity to simulate breathing. Changing the voltage to the reciprocating linear actuator 532 allows for modulating the respiratory rate within the tissue phantom. Various mixtures of CBG with other types of US compliant materials, then proceeded to tensile test on a uniaxial test platform. Stress-strain curves were generated, with pure CBG 10 having the most favorable behavior (FIG. 6), with a Young’s Modulus of 37 kPa.

[0042] For abdominal scan sites, the corresponding organs were placed to match anatomical relevance (FIG. 5C). The cavity was then filled with a 10% gelatin mixture made with a 2:1 ratio of water and evaporated milk with flour at a 1 % w / v concentration. The gelatin was poured in multiple layers to avoid displacement of the organs. The phantom was then placed at 4°C to allow the gelatin to solidify. US compatible bladders (or pouches) were used to create modular injuries like those illustrated in FIGs. 7A-7B. FIG. 7B illustrates a bladder with a set of three cavities that are independently controllable and configured to be placed between the rib cage and the lung surrogate. The bags were sealed with a vacuum sealer, and tube fittings were added to these pockets which allowed for a syringe to be connected to ensure precise control over the amount of water or air added.

[0043] Thoracic injury creation was done with fluid (HTX) or air (PTX) filled bladders being placed between the lung and thoracic phantom of variable volumes to create modular injuries for each injury status. By removing the bladder, injury negative images can be captured. US images can be captured as B-Mode or M-Mode scans for visualizing lung motion and underlying injuries for evaluation by Al image interpretation models.

[0044] Full AH injuries were created by inserting up to 40 mL of water in the injury bladders. Full HTX and PTX were created by inserting up to 40 mL of water and up to 80 mL of air,respectively. By controlling the amount of fluid or air, injury severity was able to be modulated to simulate injury progression. This feature enabled the creation of varied injury levels, ideal for training the Al models to not only detect the presence of an injury but also assess the severity of the injury from US scans.

[0045] A mechanism to simulate breathing was successfully developed, enabling the imaging of physiologically representative breathing at multiple scan sites. Due to limitations of the DC motor controller, simulated breathing rates in the prototype ranged from 24 to 84 breaths per minute (BPM), faster than typical breathing rates. US M-mode images for both negative and positive PTX and HTX thoracic conditions at various BPMs can be seen in FIG. 8.

[0046] The mechanism used to simulate injuries was used to model injury progression. Empty US sleeve pouches were first placed at each scan site, and US images of negative injury states were collected. The injury severity was then progressed by filling the bags with variable quantities of fluid or air, simulating varying levels of injury. B-mode ultrasound images for the HTX and LUQ scan points showing increasing injury volume were then captured as seen in FIG. 9.

[0047] FIGs. 5A-5C illustrate a phantom built according to the above method. FIG. 5C illustrates the proximate locations for the different scan points and abdominal cavities. The phantom 100 is ready for loading in the lung 520 attached to the sled 400 illustrated in FIG. 4 and the organs illustrated in FIG. 3. For the bladders like those illustrated in FIGs. 7A and 7B that will be filled can be placed into the phantom as well. The bladders may be incorporated into an organ or placed next to it. The bladders are designed to collapse to nothing and when filled will make space within the phantom by pushing against the organs.

[0048] FIG. 10 illustrates the front side of the phantom 100 with the pink lungs viewable. The foam 520 is aerated so that there is air present inside the foam 520 that makes it look like there is air movement (e.g., replicating the lung alveoli) as the sled 410 moves back and forth, for example at 160 reps per minute. FIG. 11 illustrates a close-up of the upper torso area, the expiration system 530, different conduits 746 running out from the phantom body 101 , and the rotation mechanism 194 of the frame. The rotation mechanism 194 includes a pair of motor- driven gears configured to rotate the frame about the axis between them.

[0049] FIGs. 4-5B illustrate the configuration of the structure to facilitate lung motion according to at least one embodiment. This approach allows for lung motion across most intercostal spaces simultaneously, allows for motion during hemothorax image capture, and allows for modular thoracic injury severities. The approach makes use of a reciprocating linear actuator configured for repetitive linear motion. The mechanism is attached to a sled 410 holding a aerate silicone foam 520 formed to the shape of the thoracic cavity to serve as a lung in the setup. In at least one embodiment, the sled 410 is U-shaped with top and bottom restraining walls 412, 414 attached to a support base 416 where top and bottom are relative to a longitudinal axis running the length of the phantom with top restraining wall 412 closer to the head and the bottomrestraining wall 414 closer to the abdomen as illustrated in FIGs. 4-5B. The top restraining wall 412 includes a connection 415 for attaching to the shaft 538. The spikes 418 present on the sled 410 attach to the foam lung 520. Movement of the sled 410 creates realistic B-mode and M- mode image capture of the thoracic cavity as the silicone foam lung 520 moves back-and-forth along the longitudinal axis within the rib cage 522. Movement of the sled 410 is the result of the expiration system 530.

[0050] The expiration system 530 includes a series of components 531-539. An actuator 532 is mounted on a support wall 533 with an actuator shaft 533 passing therethrough. The actuator 532 is configured to rotate a control arm 534 about its rotation axis passing through the actuator shaft 533. The other end of the control arm 534 connects to a linkage arm 536 to rotate the connected end to lead to linear movement of a shaft 538 connected at the other end of the linkage arm 536. The shaft 538 extends into the phantom 100 (as seen in FIG. 5C, 10, 1 1) to attach to the sled 410, e.g., at the connection mount 415 on the closest wall 414. In at least one embodiment, the shaft 538 passes through an alignment component 539. The expiration system 530 is oriented to be in line with a longitudinal axis of the phantom 100. The frequency of the actuator 532 mimics a breath rate, which was simulated between 24 and 84 BPM, which is faster than normal. In at least one further embodiment, the current range of respiratory rate can be adjusted with higher torque, lower speed actuators. The support wall 533 may have a variety of forms including the illustrated configuration, omitting the horizontal base, extending the horizontal base to the actuator side, including cooling holes passing through the vertical wall, including mounting holes in the horizontal based, and / or any combination of these configurations.

[0051] Modular injuries are possible to simulate by vacuum sealed air or fluid filled bladders 742 or a bladder with multiple chambers 744 being inserted between the ribs 522 and the lung 520 to create pneumothorax (PTX) or hemothorax (HTX) injuries, respectively. When a bladder is present, it allows for simulation of decompression in response to insertion of a needle. Next steps will combine these improved injury mechanisms with the prior abdominal injury setup (combined with improved modularity techniques) to create a robust injury phantom capable of positive and negative injuries in the abdomen, for example, by using bladders 742. Each bladder / chamber is connected to a conduit 746 to facilitate delivery and removal of fluid and / or air from the bladder / chamber either independently of each other or one conduit connecting to multiple bladders / chambers.

[0052] The lung motion aspect was accomplished by utilizing a reciprocating linear actuator 532 with a sled 410 which was resin printed, the lung phantom 520 was made from a foam placed into the sled as seen in FIGs. 5A-5B. Simulating lung motion with the expiration system 530 was key to simulating PTX and HTX injuries through modularity. It was found that the “breathing” rate could be altered through adjusting the stroke length of the expiration system 530, with a longer stroke length corresponding to a decreased breath rate and vice versa.

[0053] Examples of the material that may be used for the different organs illustrated in FIG. 3 are provided in Table 1. Casts were made for all internal organs, the bulk of the phantom was poured around the ribs, cartilage, and sternum.Table 1

[0054] The phantom was tested beginning with image capture and image processing to create a training set of images for training an Al model.

[0055] Two types of Al models were developed using US images: anatomical guidance (AG) and injury diagnostic (ID). These models were trained using a dataset of frames extracted from 10 second US video clips captured using a Vscan Air (GE Healthcare). For AG model training, two sets of US B-mode videos were collected: 1) medial to lateral and 2) distal to proximal scan swipes at each site. For abdominal scan sites, the convex probe (2-5 MHz) was used for video capture aimed at keeping anatomical landmarks of interest in view, namely bladder and kidneys. For thoracic scan sites, the linear probe (3-12 MHz) was guided along the ribs, starting from the upper chest and moving across the armpit to the lower ribs. Sets of images were captured for both negative and positive injury. Once collected, data was prepared by labeling the following anatomical landmarks with bounding boxes: kidneys for the RUQ and LUQ views, bladder for BLD view, and ribs for the thoracic views. T o achieve this, US clips were cropped to remove the US system user interface and separated into individual frames, then labeled using video annotation software, such as COCO annotator and MATLAB toolboxes.

[0056] For ID models, three B-mode US videos were collected at each scan site. For abdominal scan sites, these videos imaged the anatomical features of interest while the probe was slowly tilted to capture views from different angles. For thoracic scan sites, the probe was placed at intercostal sites, ensuring that two ribs and the inner pleural space were in view. These US videos were captured for both negative and positive injury states and then organized into therelevant ID model’s training data. To introduce subject variability, ID and AG US images were captured on three manufactured iterations of the prototype. Each iteration of the prototype was made by repositioning the organs and then repouring the gelatin mixture so that anatomical variability could be present within the collected images.

[0057] The processed US images were used to train Al models for each scan site using two approaches. First, object detection models for AG were developed for anatomical detection of ribs, kidneys, and bladder features at the respective scan points using a YOLOv8 model architecture (Yaseen, “What is YOLOv8: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector,” 2024) which had previous success in US applications (Hernandez Torres, “Real-Time Deployment of Ultrasound Image Interpretation Al Models for Emergency Medicine Triage Using a Swine Model,” Technologies, 2025;13:29) utilizing the YOLOv8-s pretrained object detection weights. Next, classification models using the same YOLO framework utilizing up to five different phantoms were used for training ID models while the diagnostic data from three separate phantoms were blind subjects used to evaluate model performance. In total, training data for AG models contained 2200-3800 images for training with an extra 550-850 images for validation data while abdominal ID model training data contained 2800-3000 images with about 300 images separated for validation and thoracic ID models were trained off a much smaller dataset of 368 images with just 30 M-mode images used for validation.

[0058] Each scan site had binary model types (positive or negative for injury) except for the lung model which was three class (HTX, PTX, or negative). Both kinds of YOLO models utilized the following default training settings: 100 training epochs, a batch size of 16, a learning rate of 0.1 , a weight decay of 0.0005, and no early stopping. Default YOLO image augmentations were also applied to a fraction of the training data, including adjusting the hue, saturation and brightness of the images, translating the image vertically or horizontally by a factor of 0.1 of the image size, scaling the image by a gain factor of 0.5, flipping the image by the vertical axis, and erasing part of the image so the model is able to recognize less obvious features in the image.

[0059] Each of the conventional YOLOv8 ID models’ performance was then compared to a classification model whose architecture was developed utilizing a Bayesian optimization process. The Bayesian optimization approach used a series of convolutional layers with ReLU activators and max pooling layers followed by a fully connected layer, dropout layer, and final classification layer. The number of CNNs, number of nodes, filter size, dropout amount, fully connected layer size were all hyperparameters as were affine augmentation types with variable extent of turning each augmentation off as additional hyperparameters. Lastly, the batch size and learning rate were set as hyperparameters in the optimization setup. The same training data for YOLOv8 training was used for this approach with the exception that one additional blind test phantom was used for validation with a validation loss patience set to 5 epochs for a maximum of 20 epochs. In summary, between 2,000-3,000 US images were used for abdominal scan sites for trainingwhile approximately 350 images were used for thoracic model training as less M-mode images were captured. An additional blind test phantom provided test performance during Bayesian optimization where the goal of the 100-iteration optimization problem was to maximize blind test accuracy. The top 5 model configurations were trained in triplicate for 100 epochs with a validation loss set to 5 epochs. An additional blind test phantom provided a way of testing performance on the optimized model. These US scans were further used to construct GradCAM to highlight regions of the US scan where Al decisions were focused to bolster explainability of the Al model results.

[0060] AG object detection models for the ribs, kidney, and bladder were successfully developed using YOLOv8 with training intersection-over-union (IOU) scores of 0.821 , 0.798, and 0.906, respectively. However, when evaluating the object detection models with blind test phantoms, performance was slightly reduced with IOU scores of 0.648, 0.516, and 0.536 for the ribs, kidney, and bladder, respectively.

[0061] For diagnostic Al, overall performance metrics with blind test phantoms for YOLOv8 and highest performing Bayesian optimized models are summarized in Table 2. YOLOv8 outperformed the Bayesian models for the thoracic with approximately 0.10 higher accuracy, but trailed behind the Bayesian model results for the other scan sites. Most notably were RUQ and BLD results where accuracy was 0.230 and 0.323 higher with the Bayesian model compared to YOLOv8, respectively. GradCAM overlays for negative (top row) and positive (bottom row) image types for each scan site with the Bayesian Optimized model highlights the model's capability to identify key anatomical features in the case of RUQ, BLD, and LUQ scan sites as depicted in FIG. 12 (representative images). The fluid accumulation site is identified in the image to exemplify how the Al model identified these sites. For the m-mode lungs image, the Al model identified breathing signs in the negative image, fluid accumulation in the pleural space for HTX, and steady pleural line for the PTX site (FIG. 12), indicating successful model training for recognizing relevant anatomical patterns.Table 2

[0062] Overall, the phantom was successful at creating realistic US images and showed positive results in training Al models. In this work, an US compatible thoracic phantom with modular injuries and the capability for simulated lung motion was developed. The drive system allowed for the collection of realistic lung images and the simulation of PTX and HTX injuries, both of which were integral for training the Al models in this study.

[0063] The features of this embodiment were able to overcome the shortcomings of other commercially available phantoms. The ability to alter injury severity, image lung motion along the entire thoracic region, and adjust respiratory rates are integral features for developing an effective Al that can be viable at the point of injury on a battlefield setting. The phantom may be used not only as an Al training tool but also as a resource for medical personnel training. Ballistic phantoms have been used to educate trainees in conducting eFAST examinations (Amini, “A novel and inexpensive ballistic gel phantom for ultrasound training,” World J Emerg Med., 2015;6:225-8.), due to their durability and the realistic US images they generate. During training scenarios, a phantom capable of producing image variance will expose trainees to varied injury severities, thus providing valuable experience to triaging personnel.

[0064] In this study, the use of the phantom in Al model development was highlighted. The most important capability of the phantom for this task is providing sufficient anatomical features along with subject variability to enable Al models to perform effectively on blind test subjects while relying on relevant anatomical features applicable to real-life scenarios. The object detection AG models performed well during training and maintained blind performance higher than a 0.50 IOU performance threshold. However, there was a noticeable drop in performance between training and blind testing, indicating the models may be overfit to the training data. Therefore, greater subject variability or enhanced image augmentation may be required to improve AG model performance. Successful object detection AG models can aid in real-time US scans to confirm to the end user that the US probe is properly positioned at the scan site, lowering the skill threshold for this triage examination if properly implemented.

[0065] In addition, diagnostic models were developed with mixed success. YOLOv8 models were initially developed; however, performance was insufficient during blind testing due to overfitting issues for most of the abdominal scan sites. Instead, how the phantom could be used to tune Bayesian optimization parameters to customize model performance for this specific usecase was highlighted. All scan site blind test exceeded 70% accuracy, and GradCAM overlays illustrated that Al models identified anatomical injury landmarks, a critical design feature required for the developed phantom to be suitable for this Al application.

[0066] Additionally, the injury diagnostic models showed promise in their ability to evaluate injury status using relevant anatomical features. Model performance may be improved through further refinement and optimization with approaches that focus on model generalization. Overall, the phantom has the potential to serve as an effective subject for developing and testing Al models for US image interpretation and overcoming the challenges faced in the deployment of US imaging techniques in combat casualty care settings.

[0067] In another phantom embodiment, the tissue phantom uses a synthetic gelatin component mixed with Talc powder to create a realistic US tissue. After construction using a method similarto that illustrated in FIG. 2, cavities were carved out where the bladder, right upper quadrant, left upper quadrant and thoracic scan sites are present. 3D organs and ribs were printed and used to create reverse cast molds using silicone epoxies. These were used to make ballistic gelatin organs which were arranged together anatomically with or without hemorrhage pockets present, followed by suspension in food-grade gelatin mixed with flour within the carved- out tissue phantom cavities. For thoracic use, a spinning foam wheel was constructed which could be held against various intercostal spaces for creating realistic lung motion as obvious in negative eFAST thoracic images. However, HTX injuries were only possible in a static setup not with lung motion as is more physiologically relevant. Further, the thoracic lung motion approach could only be applied to a single intercostal space at a time, limiting its utility.

[0068] A full torso phantom was developed and included of the main internal organs that are ultrasonically viewed when performing an eFAST exam. An open access repository containing ready-to-print 3D datafiles was used to obtain models for: the bottom lobe of both lungs, liver, spleen, stomach, both kidneys, bladder, rectum, ribs, costal cartilage, and sternum (Mitsuhashi 2009). Table 1 above summarizes the material and 3D-printing method used for each part. Fused deposition modelling printing was performed using a Raise3D Pro2 Plus printer (Raise3D, Irvine, CA, U.S.), and stereolithographic printing was done using FormLabs’ Form2 or Form3L printers (FormLabs, Somerville, MA, U.S.).

[0069] Since the 3D-printed plastic organs have relatively low melting points, a multi-step casting process was used. Each organ was first 3D-printed, and then cast with Dragon Skin 10 NV (Smooth-On, Macungie, PA, U.S.). The Dragon Skin’s heat resistant properties made it ideal for use with CBG (Clear Ballistics, Greenville, SC, U.S.). Sufficient volumes (varied by organ) of CBG were melted at 130°C and mixed with talc powder (Fasco Epoxies, Ft. Pierce, FL, U.S.) at 0.25% (w / v) concentration to pour each organ mold. Specifically for the stomach, lungs, and rectum models, this same CBG mixture was vigorously stirred before pouring to mix in anddistribute small bubbles which mimic the natural state of these air-filled organs. Conversely, two different bladder sizes were poured with pure CBG without talc powder, as it is a fluid-filled organ.

[0070] Once the ribs, cartilage, and sternum were 3D-printed they were assembled using custom brackets. A schematic of the full assembly generated using Solidworks CAD (Waltham, MA, U.S.) with each anatomical component identified is shown in FIG. 13. In place of the backbone, silicone tubing with notches was used to support the posterior ends of the ribs. To protect the inside of the thoracic cavity, rubber sheets were sown together and attached to the ribs with wires, preventing CBG from leaking into the cavity when pouring the bulk of the torso.

[0071] A torso cast was made with EpoxACast HT670 (Smooth On, Macungie, PA, U.S.) using a plastic mannequin (Amazon, Seattle, WA, U.S.). The mannequin was modeled after a human male, from the upper thigh region to just above the clavicle and placed supine into the EpoxACast HT670 resin mixture. A rim was fixed normal to the coronal plane and attached around the perimeter of the mannequin allowing the front contour shape to be captured while adding depth to the cast as opposed to fully submerging the mannequin model. This permitted space to position the internal components within the final phantom mold. The bulk of the torso was made of the same CBG with 0.25% (w / v) talc powder mixture. The ribs were placed in the empty cast, lined with rubber sheets, and then layers of liquid CBG mixture were poured until the ribs were covered. Layering was used to allow bubbles to escape after each portion was poured, ensuring minimal bubbles were trapped in the final phantom. The bulk of the phantom was allowed to cool and fully solidify overnight.

[0072] After the bulk was ready, the area right behind the ribs was carefully carved to remove the rubber sheets and have access to insert the internal organs. All the carving for organ placement was performed from the back to preserve the front and sides of the torso for US imaging. The cavity inside the ribs was expanded to the waistline, allowing for enough space to fit the stomach, liver, kidneys, lungs, diaphragm, and spleen. To allow superior access to the thoracic cavity, any neck tissue was removed, leaving the thoracic cavity open from the superior and posterior sides of the phantom. A second cavity, in the pelvic area, was also carved from the back until there was approximately 2 cm to the surface (front of phantom). This cavity was big enough to fit both bladder sizes along with the rectum. Carving for all areas was performed with a scalpel blade, and a hot knife (Modifi3D, Coalville, UK). While carving the bulk of the phantom, it was left in the original cast for ease of mobility.

[0073] US imaging was performed using three different US systems: Sonosite Edge (Fujifilm Sonosite, Bothell, WA, U.S.), Sonosite PX (Bothell, WA, U.S.), and Terason 3200t (Terason, Burlington, MA, U.S.). For the chest scan points a linear-array probe from each system was used to collect M-mode images. All other views were scanned with curvilinear and phased-array probes to obtain a 30 second B-mode clip of the scan points. Additional details specific to each scan point are described in the sections below. A polyurethane quick set foam resin (McMaster-Carr, Elmhurst, IL, U.S.) platform was created to the same dimension as the phantom cast to act as a base for the phantom to lay on its back during US imaging.

[0074] PTX baseline images were created by rotating a 2 cm wide and 1 . 1 cm thick foam ring (i.e., lung segment) under an intercostal space in a thoracic area of the phantom. The foam piece was attached with a cyanoacrylate adhesive (Loctite, Dusseldorf, Germany) around an acrylic disc with a diameter of 6.25 cm. The disc was then attached to the end of the rod and placed under the intercostal space region of interest illustrated in FIG. 14. The rotating motion was created by an actuator 1432 (Amazon) powered by a 12 V power supply at 12 revolutions per minute. The motion was transferred to inside the phantom through the neck cavity using a hexrod 1434. The t-rail containing the actuator 1432 was angled (see also FIG. 13) to allow complete contact between the intercostal space and the lung segment 1420. The angle is relative to the longitudinal axis passing through the phantom. Twelve evenly spaced cuts were made to the foam midpoint to give the M-mode images more granularity. Prior to imaging, US gel (Aquasonics Enterprises, Gary, IN, U.S.) was applied to the lung segment. Rotating the device resulted in baseline images as lung motion is mimicked by rotation. PTX positive images were collected by removing the device from view. A total of four M-mode images for each condition, negative and positive PTX, were captured with the three US systems.

[0075] The bladder and rectum CBG models were attached to each other using melted CBG and placed inside the pelvic cavity. A 10% gelatin mixture dissolved in water and evaporated milk, with flour at 0.25% w / v concentration was added around the organs to fill out the space in the cavity. This mixture has been previously used to simulate tissue in US phantoms (Hernandez- Torres, (2022), “Using an ultrasound tissue phantom model for hybrid training of deep learning models for shrapnel detection,” J. Imaging 8, 270). The gelatin mixture was left to solidify at 4°C for approximately 2 hours covered in plastic wrap to prevent gelatin dehydration. To create positive AH injuries, a CBG hypoechoic pocket was cast in a custom-made 3D mold and attached between the bladder and rectum. Once gelatin solidified, the phantom was placed supine on the foam base for US imaging.

[0076] A wall was created inside the rib cavity to separate areas designated to PTX and HTX models. The wall was created by placing two laser cut acrylic panels inside the cavity with CBG poured on top to create a fluid seal. For AH and HTX models, a CBG diaphragm model was developed by spreading melted CBG mixed with 0.25% (w / v) flour into a thin layer. The diaphragm was then attached to the lungs and to the ribs in the thoracic cavity of the phantom using melted CBG. The liver, spleen, stomach, and kidneys were attached to each other using melted CBG as pasting material in their approximate anatomical locations. The CBG organs were then placed inside the rib cavity. A total of 3L of the 10% gelatin mixture simulating tissue was used to fill the cavity space around the organs. The phantom was held at 4°C to allow the gelatin mixture to solidify. T o create a positive HTX injury, a sheet of CBG without any air bubbleswas placed in between the ribs and lungs. For positive AH injuries, the hypoechoic pocket used for the pelvic view was also attached in between the liver and kidney or spleen and kidney for RUQ and LUQ AH, respectively. Imaging was conducted with the phantom remaining prone in its cast to maintain structural integrity.

[0077] A commercial eFAST trainer (Simulab, Seattle, WA, U.S.) was used for scan point view location confirmation and US image comparison. The trainer included two normal patients and three patients with different injury combinations. As the probe goes near the different scan points the software shows representative images or videos of the region. Some of the abdominal scan points offer longitudinal and transverse views, as well as videos. The lung scan points have both B-mode and M-mode images. US scans from the software were screen-recorded and framed, using the procedure described below and shown as representative human images throughout the result sections, for comparison.

[0078] All US images and clips captured were named according to the US system they were recorded with, injury type, injury severity, and probe used to scan. For the PTX view, images were split into baseline and positive folders and then processed using MATLAB R2022b (MathWorks, Natick, MA, U.S.). The images were cropped to remove the user interface of each US system, leaving only the M-mode region of interest. For cropping, pixel coordinates of the upper corner were identified, as well as the length and width of the window of interest. To boost the number of US images, a 4-pixel rolling window was used for further cropping, yielding 108 image segments per M-mode image, following the process described previously (Boice et al.). The image segments were then resized to 512*512x3 using MATLAB batch image processing.

[0079] For all the other scan points, frames were extracted from each 30 second clips using ffmpeg via a Ruby script. Images for the pelvic view were separated into baseline and AH positive folders, for binary classification. Images for the RUQ and LUQ were split into three categories: baseline, AH positive, and HTX positive. At this point images were processed with MATLAB to crop and remove the patient information and user interface of each US system. Similar approaches were used to determine the top left corner, width, and height pixel coordinates used for cropping. The images were then resized to 512x512x3.

[0080] A previously developed deep learning architecture, ShrapML, was tuned for image classification of US images (Snider, (2022), “An image classification deep-learning algorithm for shrapnel detection from ultrasound images,” Sci. Rep. 12, 8427). Briefly, the architecture is an original, Bayesian optimized, convolutional neural network with 6 convolutional and two fully connected layers which uses a RMSprop optimization function with 430k trainable parameters. This architecture has been used to develop a PTX detection model (Boice et al.), which was used to test the images obtained from the full torso tissue phantom for the PTX scan points. The same algorithm architecture was used to train new models for the other three scan points. Prior to training, images were augmented using X-Reflection, Y-Reflection, X-Translation (-60 to 60pixels), Y-Translation (-60 to 60 pixels), and Image Rotation (-180 to 180°). Example images after augmentation are shown in FIG. 15.

[0081] Training of the eFAST models was conducted using MATLAB. Phantom images were processed according to the injury type and were split into 70%, 10%, and 20% for training, validation, and testing, respectively. The models were trained for up to 100 epochs with a validation patience of 5, learning rate of 0.001 , and a batch size of 32. Training was performed using either an Asus ROG Strix running Windows 11 , 12th gen 14 core i9-12900H (2500 MHz), 16GB RAM, and a NVIDIA GeForce RTX 3070 Ti (8 GB VRAM) or a Lenovo Legion 7 running Windows 11 , AMD Ryzen 9 5900HX (3300 MHz), 32GB RAM, and a NVIDIA GeForce RTX 3080 (16 GB VRAM).

[0082] Performance of Al models was evaluated on blind, holdout testing images not used during training. Confusion matrices were created to classify testing performance as true positive (TP, correct injury identification), true negative (TN, correct baseline identification), false positive (FP, injury identified when not present) or false negative (FN, injury not identified when present) depending on the accuracy of the prediction to the ground-truth label. These labels were used to calculate accuracy, precision, recall, specificity, and F1 score using widely accepted calculation approaches (Shri Varsheni, 2021), and to create confusion matrices for each model using GraphPad Prism (San Diego, CA, U.S.). In addition, area under the receiver operating characteristic curve (AUROC) was calculated for each label. For the PTX view, eight previously developed replicate Al models were used to make blind test predictions. Performance metrics were calculated on blind test images for three trained models for RUQ, LUQ, and pelvic Al models, and averages and standard deviations were calculated for each metric. Images from three US machines were merged and treated as a single training set unless otherwise specified. In addition to performance metrics, GradCAM was used to assess what features in an US image was driving the Al model prediction (Selvaraju 2017). GradCAM heat map overlays were created using MATLAB R2022b for a subset of test images. “Hot spots” in the overlay correspond to high weighted regions, driving the model prediction for the respective US image.

[0083] Diagnosis of PTX using US has key landmarks in both B-mode and M-mode US imaging. While US scanning the full torso phantom with the actuator running, B-lines and sliding lung were present in real-time, characteristic of healthy breathing lungs. FIGs. 16A-16E illustrate a recreation of a PTX injury in the phantom. When scanning in M-mode, the “seashore” sign became apparent (FIG. 16A) while the actuator was running and then converted to a “barcode” sign (FIG. 16B), when the simulated lung was no longer in contact with the pleura. The same patterns can be observed in the commercial eFAST trainer (FIGs. 16C, 16D).

[0084] Previously trained Al models, successful at detecting PTX in swine and simple tissue phantom M-mode images, were used to make predictions for the image sets collected in the full torso phantom (Boice et al.). Initial test results had unsatisfactory performance, leading to theseparation of data for each US imaging system to provide a more granular understanding of the results. FIGs. 17A-17D illustrate Confusion Matrices for PTX model predictions on test M-mode images collected from the phantom. The Sonosite Edge resulted in all the predictive PTX models (n=8) classifying the images as negative for PTX, resulting in 50% accuracy (FIG. 17A). However, this trend did not continue for the other US machines, with Sonosite PX having similar rates for both false positive and false negative outcomes and Terason being biased toward false negative results (FIGs. 17B, 17C). The accuracy for both systems was 85%-87% (Table 3). On average, across all the US systems, results were heavily skewed towards a false negative outcome (FIG. 17D) due to the Sonosite Edge predictions and had an overall accuracy of 74%.TABLE 3Table 3 shows performance metrics for pneumothorax model prediction on test M-mode images collected using the developed tissue phantom. Results are shown for individual US systems and as an average across the three systems. Performance metrics were averaged across n=8 previously trained pneumothorax models and standard r these replicates were calculated.

[0085] To further evaluate differences in predictions, GradCAM overlay masks were generated for test images from each US system. FIG. 18 illustrates GradCAM overlays for PTX Al model predictions on test images collected using the phantom. Representative US M-mode scan segments for each US system are shown without and with the GradCAM overlay for (A) PTX negative and (B) PTX positive images. Areas of high importance to the Al model as determined by GradCAM are indicated as red-yellow while lower importance regions are denoted as green-blue hues. Similar, features were being tracked in negative or baseline images for all US systems, with the heat map being focused on the “seashore” sign in the M-mode image. For PTX positive images, all the Sonosite Edge images failed to detect or track any feature, while the majority of predictions from Terason 3200t and Sonosite PX tracked the “barcode” sign on the M-mode image segments. This example highlights the variability that US systems can have on Al predictions and the relevance of including these effects in this stage of model development.

[0086] FIGs. 19A-19E illustrate a pelvic view in the phantom. Abdominal hemorrhage diagnosed in the pelvic view looks at blood pooling behind the bladder towards the rectum in FIG. 19A. In the tissue phantom, this was replicated by placing a hypoechoic pocket betweenthe bladder and rectum in FIG. 19C. The pelvic view in the full torso phantom for both baseline and AH in FIGs. 19B, 19C, resembles the US images from the commercial trainer for the same views and injury in FIGs. 19D, 19E, except for US image depth and contrast.

[0087] Deep learning predictive models for the pelvic view were trained using tissue phantom images as Al models had not been previously developed for this imaging application. A subset of images was used for training (70%) and validation (10%) and included images from three US systems to help overcome prediction biases. FIGs. 20A-20C illustrate performance results for an Al classification model trained for the pelvic eFAST view. Using blind test images (20%) from all US systems, the resulting trained Al model successfully identified true positive and negative images with a low false positive and negative rate as illustrated in FIG. 20A. Evaluating Grad- CAM predictive overlays, the Al was tracking the region below or around the bladder for positive and negative AH predictions which correlates with where fluid would be found as illustrated in FIGs. 20B, 20C. Triplicate deep learning models were trained with a similar accuracy, 99% on average as shown in Table 4.TABLE 4Table 4 shows a summary of performance metrics for the pelvic Al model. Average results and standard deviations are shown for n=3 trained models. ROC (receiver operating characteristic) Curve.

[0088] The abdominal scan point on the right side of the abdomen can look for HTX or AH. When the US image is shallower and focused on the posterior, lower rib area HTX can be diagnosed by blood accumulating within the pleural space. FIGs. 21A-21D illustrate a RUQ view in the phantom. In the developed phantom, this was simulated by placing a completely clear sheet of ballistic gel as illustrated in FIG. 21A between the ribs and bubbly lung, generating a HTX positive image in the phantom shown in FIGs. 21 B, 21 C. US scan focused on the liver and kidney looks at the region known as Morrison’s pouch for any signs of abnormality. Here, differences are noticed if blood is pooling between the liver and kidney shown in FIG. 21 E, as placed in the tissue phantom shown in FIGs. 21 F, 21 G. Images were similar to when comparing to the same anatomical location in the commercial eFAST trainer human images shown in FIGs. 21 D, 21 H.

[0089] Similar to the pelvic view, an Al model was trained for this specific application using the previously developed ShrapML framework. However, the model was modified to allow for three categorical outcomes - AH, HTX, and negative for either injury. FIGs. 22A-22D illustrate performance results for an Al classification model trained for the RUQ eFAST view. A model was successfully trained using US images from all three US systems, with blind image predictions identifying each of three categories with high accuracy and without an obvious bias toward any false category illustrated in FIG. 22A. Evaluating Grad-CAM overlays, AH-positive predictions were often focused on the area around the hemorrhage region shown in FIG. 22B and HTX- positive predictions were focused on the middle of the US scan where the thorax hemorrhage effects were most obvious shown in FIG. 22C. Negative predictions had less of a consistent focus, with most heat map attention on the organs at the center of the US scan shown in FIG. 22D. Overall testing accuracies for each of the three categories were 98.6%, 98.7%, and 97.6% for AH, HTX, and negative, respectively, in Table 5.TABLE 5Table 5 shows a summary of performance metrics for right upper quadrant Al model. Average results and standard deviations are shown for n=3 trained models for each classification category: abdominal hemorrhage positive, hemothorax positive, and negative for both injuries.aArea Under the ROC Curve.

[0090] Similar to the RUQ view, the LUQ scan focuses on the upper left side of the abdomen, where HTX and AH can be diagnosed as well. FIGs. 23A-23H illustrate a LUQ view in the phantom. For HTX, the US diagnosis focuses on fluid accumulating between the pleural spaces shown in FIG. 23A and shows as a dark fluid band superficial to the lung, visible through an intercostal space shown in FIGs. 23B, 23C in the full torso phantom. The AH on the left side of the body can be diagnosed by a dark hypoechoic strip of blood between the kidney and spleen shown in FIGs. 23E, 23F, as arranged in the US tissue phantom in FIG. 23D. When compared to human US scans in FIGs. 23G, 23H, images captured with the tissue phantom were anatomically similar.

[0091] An approach similar to the RUQ view was taken to train a three category — AH positive, HTX positive, and negative for both injuries — classification model for the LUQ views. FIGs. I SA-13D illustrate performance results for an Al classification model trained for the LUQ eFAST view. Resulting models had a strong affinity towards true positive and true negative predictions across the three categories shown in FIG. 24A. The AH positive predictions were often tracking the image region where blood was present shown in FIG. 24B and HTX predictions continued this trend of tracking the hemorrhage region but sometimes detecting the darker area of the overall image shown in FIG. 24C. Negative image predictions were mostly focused on a cross-section of the image and the absence of hemorrhage in those regions shown in FIG. 24D. In summary, model accuracy for the LUQ view for the test set by category was 97.5%, 98.0%, and 98.4% for the AH-positive, HTX-positive, and negative categories, respectively, in Table 6.TABLE 6Table 6 shows a summary of performance metrics for left upper quadrant Al model. Average results and standard deviations are shown for n=3 trained models for each classification category: abdominal hemorrhage positive, hemothorax positive, and negative for both injuries.aArea Under the ROC Curve.

[0092] There are some limitations with this phantom and trained Al models that should be noted. First, the first phantom is slow to create compared to using commercially available trainers. The bulk of the tissue phantom housing the organs only needs to be made once, but the gelatin embedded US imaging sites must be newly cast each time which can take approximately 2 hours to solidify prior to imaging. Second, while the phantom contains some subject variability, it pales in comparison to the variability expected to be seen in a human subject population. The repouring of the US scan regions can assist with this, but organ sizes and hemorrhage severity were not varied in this work, with the exception of the fullness of the bladder. Third, the PTX negative lung motion can only be viewed at a single rib space; the mechanism also does not allow for lung point generation in which partial PTX positive and negative views are evident at a single intercostal space, a phenomenon clinically used to identify PTX (Skulec 2021). Lastly, the Al models developed using the phantom need to be validated and transfer- learned with human or animal images before they are suitable for use beyond this platform.

[0093] In a further embodiment to the above embodiments, the phantom will include an external skin layer with recognizable features, such as nipples, to improve its use with computervision applications. Other improvements include more noise in organ size, placement, and injury severity to expand on the robustness of the Al training potential for the phantom.

[0094] The first phantom embodiment was developed based on the second phantom embodiment with some functional improvements: 1) a new thoracic breathing mechanism, 2) fluid injuries at each site, 3) modular injury creation, and 4) an inversion mold. To assist with the handling and preservation of the phantom, a rotating mechanism to invert the phantom during the pouring and imaging processes was developed. This was created by attaching the phantom mold to a wooden table containing a cutout large enough to fit the mold and allow tolerance while rotating. The rotating mechanism was driven by attaching two motor-driven gears to the mold. Once rotated, the mold was securely anchored using linear actuators. The first phantom embodiment allowed for modular injuries and the ability to scan the entire thoracic region allowing for simulation of the PTX / HTX injury not possible with the second phantom embodiment.

[0095] The example and alternative embodiments described above may be combined in a variety of ways with each other without departing from the invention.

[0096] As used above “substantially,” “generally,” and other words of degree are relative modifiers intended to indicate permissible variation from the characteristic so modified. It is not intended to be limited to the absolute value or characteristic which it modifies but rather possessing more of the physical or functional characteristic than its opposite, and preferably, approaching or approximating such a physical or functional characteristic.

[0097] The foregoing description describes different components of embodiments being “connected” to other components. These connections include physical connections, fluid connections, magnetic connections, flux connections, and other types of connections capable of transmitting and sensing physical phenomena between the components.

[0098] The foregoing description describes different components of embodiments being “in fluid communication” to other components. “In fluid communication” includes the ability for fluid to travel from one component / chamber to another component / chamber.

[0099] Although the present invention has been described in terms of particular embodiments, it is not limited to those embodiments. Alternative embodiments, examples, and modifications which would still be encompassed by the invention may be made by those skilled in the art, particularly in light of the foregoing teachings.

[0100] Those skilled in the art will appreciate that various adaptations and modifications of the embodiments described above can be configured without departing from the scope and spirit of the invention. Therefore, it is to be understood that, within the scope of the appended claims, the invention may be practiced other than as specifically described herein.

[0101] The claims dependent claims may be written in multiple dependent form where appropriate and for the multiple dependent claims, those claims can depend from all previous dependent claims for that independent claim.

Claims

IN THE CLAIMS:

1. A method of making a phantom comprising: pouring ballistics gel into a cast mold made from a silhouette, once an initial layer of ballistic gel was poured, placing a resin printed rib cage with a solid cavity into the mold and the remaining ballistic gel is poured in to completely fill the cast mold, removing the inner cavity, carving an access hole in the head / neck region, optionally inserting a plurality of organs into the phantom body after carving out cavities for the organs, optionally connecting a lung phantom having a sled into the rib cage, and optionally attaching the sled to a reciprocating linear actuator.

2. The method according to claim 1 , further comprising placing at least one fluid or air bladder configured to be fillable between the lung and a thoracic area of the phantom to create modular injuries and / or at least one bladder proximate one organ.

3. The method according to claim 2, further comprising causing movement of the lung phantom by activation of the reciprocating linear actuator.

4. The method according to claim 3, wherein the movement of the lung is along a longitudinal axis.

5. The method according to any one of claims 1-4, wherein the lung is made of foam.

6. The method according to any one of claims 1-4, further comprising printing the sled out of resin.

7. The method according to any one of claims 1-4, further comprising for each organ, pouring ballistic gel into a mold for each organ.

8. A phantom comprising: a body made of ballistics gel, optionally said body is in the form of a human torso; a rib cage made of resin inside of said body; a sled within a cavity of said rib cage, said sled optionally having a base on which two opposed walls extend away with one wall having a connection mount; a lung made of foam inside of said rib cage and attached to said sled, optionally said lung sits on said optional base and between said optional walls of said sled, optionally said foam is aerated to mimic breathing motion during operation; and an expiration system connected to said sled and said expiration system configured to cause movement of the sled where adjustment of the frequency leads to an adjustment in breath rate with a slower frequency corresponding to a decreased breath rate, optionally said expiration system connects to said optional connection mount.

9. The phantom according to claim 8, wherein said expiration system includes a reciprocating linear actuator connected to said lung.

10. The phantom according to claim 8, wherein said expiration system includes a support structure having a vertical wall, a reciprocating linear actuator mounted on said support structure and connected to said lung, a control arm mounted on an actuator shaft of said actuator to rotate about said actuator shaft, a linkage arm connected to said control arm, and a shaft connected to said linkage arm and said sled and configured to provide longitudinal movement of said sled in said rib cage.

11. The phantom according to claim 8, wherein said expiration system includes a rotary actuator connected to said lung.

12. The phantom according to claim 8, wherein said expiration system includes a rotary actuator attached to a drive shaft, optionally a hex member, said drive shaft is attached to said lung.

13. The phantom according to any one of claims 8-12, further comprising at least one fluid fill bladder placed between the lung and the body with variable volume to replicate different injury statuses.

14. The phantom according to any of claims 8-12, further comprising at least one air fill bladder placed between the lung and a thoracic area of the body with variable volume to replicate different injury statuses.

15. The phantom according to any one of claims 8-12, further comprising a plurality of organs.

16. The phantom according to claim 15, wherein each organ is made of ballistic gel or naturally dried gelatin.

17. The phantom according to claim 15, wherein the organs are selected from a diaphragm, a spleen, a liver, a stomach, a kidney, a bladder, a rectum, and an intestine.

18. The phantom according to claim 15, wherein the organs include a spleen, a liver, a kidney, and a bladder.

19. The phantom according to claim 15, further comprising at least one organ bladder proximate to at least one organ.

20. The phantom according to claim 15, wherein said organ bladder is fillable with air or fluid to press up against the at least one organ that it is near.21 . A method of making a phantom comprising: pouring ballistics gel into a cast mold made from a silhouette, once an initial layer of ballistic gel was poured, placing a resin printed rib cage with a solid cavity into the cast mold and the remaining gel is poured to completely fill the foam cast, removing the inner cavity, carving an access hole in the head / neck region,optionally inserting a plurality of organs including a lung segment into the phantom body, and attaching the lung segment via a shaft to an actuator.

22. The method according to claim 21 , further comprising placing at least one fluid or airfilled bladder between the lung segment and a thoracic area of the phantom to create modular injuries.

23. The method according to claim 21 , further comprising causing rotary movement of the lung segment by activation of the rotary actuator.

24. The method according to any one of claims 21-23, wherein the lung segment is made of foam.

25. The method according to any one of claims 21-23, further comprising for each organ, pouring ballistic gel into a mold for each organ.

26. The method according to 21-23, wherein the actuator is a rotary actuator or a reciprocating actuator.

27. A phantom comprising: a body made of ballistics gel, optionally said body is in a form of a human torso, a rib cage made of resin inside of said body, a lung segment made of foam inside of said rib cage, and an expiration system having an actuator connected to said lung segment and said actuator configured to cause movement of the lung segment.

28. The phantom according to claim 27, further comprising at least one fluid fill bladder placed between the lung segment and a thoracic area of the body with variable volume to replicate different injury statuses.

29. The phantom according to claim 27, further comprising at least one air fill bladder placed between the lung segment and a thoracic area of the body with variable volume to replicate different injury statuses.

30. The phantom according to any one of claims 27-29, further comprising a plurality of organs.

31. The phantom according to claim 30, wherein each organ is made of ballistic gel or naturally dried gelatin.

32. The phantom according to claim 30, wherein the organs are selected from a diaphragm, a spleen, a liver, a stomach, a kidney, a bladder, a rectum, and an intestine.

33. The phantom according to claim 30, wherein the organs include a spleen, a liver, a kidney, and a bladder.

34. The phantom according to any one of claims 27-29, wherein the actuator is configured to rotate said lung segment at an intercostal space to be scanned by an ultrasound probe and to simulate breathing.

35. The phantom according to 27-29, wherein the actuator is a rotary actuator or a reciprocating actuator.

36. The phantom according to any one of claims 27-29, wherein said actuator is a reciprocating linear actuator.

37. The phantom according to claim 36, wherein said actuator is part of an expiration system, said expiration system further including a control arm mounted on an actuator shaft of said actuator to rotate about said actuator shaft, a linkage arm connected to said control arm, and a shaft connected to said linkage arm and said lung segment, and configured to provide longitudinal movement of said lung segment in said rib cage38. The phantom according to any one of claims 27-29, wherein said actuator is a rotary actuator.

Images