SwiftMR FAQ PDF

🔤 SwiftMR FAQ Introduction Coverage MRI vendors and models MRI types Compatibility with other acceleration imaging techniques Body parts and images Regulatory clearance Korea MFDS USA FDA Other countries...

🔤 SwiftMR FAQ Introduction Coverage MRI vendors and models MRI types Compatibility with other acceleration imaging techniques Body parts and images Regulatory clearance Korea MFDS USA FDA Other countries IQ & Performance Scan time reduction Accuracy & Reliability Missing information Artifacts Preferences Deep Learning Workflow & Installation Workflow Installation SwiftMR FAQ 1  Cloud Competitive landscape Customer support General Liabilities Introduction 1. What is SwiftMR? a. SwiftMR is a deep learning based MR image enhancement software which enables scan time reduction by up to 50%. 2. How is that possible? a. You can already adjust some parameters on your existing scanner to reduce the scan time. But the reason they don’t use this in the clinical setting is because it results in an output of low quality image. SwiftMR can take this low quality image, process it with deep learning, and return a high quality image equivalent or better than the original. 3. What value does it provide? a. It resolves the conventional bottleneck in MR scan time and improves patient throughput. b. Compared to other modalities where the exam can be done in a few minutes, an MRI exam can take on average 30-40 minutes and at max 90 minutes. It is very hard for patients to stay still in the noisy, narrow scanner for that long, especially for claustrophobic, pediatric or elderly patients. c. For physicians, in order to diagnose and prescribe to patients in time an MRI exam is needed upfront. However, the wait time is a couple of weeks on average for tertiary hospitals. This could delay the whole treatment process. d. For the institution, an MRI scanner costs 1-2 million dollars but you can only scan 8-10 patients per day with a scanner. This limits your ROI on the machine. With just a software upgrade, you can boost the productivity of your existing scanner by 30% on average. SwiftMR FAQ 2  Coverage MRI vendors and models 1. Can it be used with all MRI vendors? a. Since it’s DICOM based, it is compatible with all MRI and PACS vendors. However, in terms of performance, consistent performance level is guaranteed for major vendors as our training dataset is comprised mainly of GSP. We have recently confirmed that it works well for Canon Toshiba and Fujifilm. 2. Can it be used with old scanners? a. Sure, you’ll see most drastic improvements in image quality with those. 3. Does that mean less drastic with newest models? a. Relatively yes. But even with latest scanners, we see customers unsatisfied with certain images. You can selectively use SwiftMR to improve those images. MRI types 1. Can it be used with both 1.5T and 3.0T? a. Yes. We have some customers saying that SwiftMR takes 1.5T to 2.0T. 2. Can it be used with low field MRIs like 0.3T, 0.5T or ultra high field like 7T? a. Unfortunately, no. 3. Can it be used with open MRIs? a. Is it low field? - If low field, cannot support. Minimum is 1.5T for now. b. However, if you can send through some anonymized sample DICOM images, we would be happy to try them out and let you know. 4. Do you plan to expand to these other field strengths? a. Maybe in the future. Our product strategy is aligned with the market share. That is, we are expanding our product coverage in the order of market demand. Eventually we might get there. SwiftMR FAQ 3  Compatibility with other acceleration imaging techniques 1. What if I’m already using acceleration imaging techniques like CS? Can I still use SwiftMR? a. Yes, SwiftMR works on top of any conventional techniques like GRAPPA, SENSE, or CS. Scan time and/or image quality will be improved from thereon. b. However, you cannot use SwiftMR on top of other DL solutions like GE AIR DL. They will crash with each other. Instead you can choose to selectively use a certain solution for certain images if you want to try out both. Body parts and images 1. Which body parts does it support? a. SwiftMR supports all body parts except abdomen and cardiac. b. Abdomen is currently under development, and we welcome research collaboration partners. 2. Which images does it support? 1. Pretty much all you need for a routine exam - T1, T2, FLAIR, PD, T2*, CE T1, CE FLAIR. SwiftMR FAQ 4  2. Brain MRA (TOF) : Image enhancement is done on the source TOF image. MIP post-processing also supported. 3. For Spine and MSK, with and without fat suppression supported. 4. SWI available. 5. DWI not yet. 6. 3D T1 for GSP. Reference materials Time-of-Flight (TOF) MRA 2D vs 3D MRA Fat suppression methods SWI DWI 3. Does it support specific protocols such as vessel wall imaging, brain metastasis imaging, TLE imaging? a. It would depend on what images(sequences) make up those orders. b. A typical vessel wall imaging protocol would be comprised of 3D PDWI(SPACE), 3D T1WI(SPACE), CE 3D T1W1(SPACE). These images are all supported. Regulatory clearance Korea MFDS 1. Is it MFDS cleared? a. Yes, it is MFDS cleared as Class I since Feb 2021. b. There are no regulatory limitations to body parts, image/sequence or parameter coverage in Korea, hence we can provide the latest updates to our home market at the earliest. SwiftMR FAQ 5  USA FDA 1. Is it FDA cleared? a. Yes, it is FDA 510k cleared (Class II) since Oct 2021. b. There are currently regulatory limitations on coverage. For US, clearance is for Brain MRI/MRA, spine, shoulder, knee, ankle, hip of non-contrast enhanced images. USA coverage c. We are filing for another submission by the end of this year to expand coverage to all body parts and all vendors. It should be cleared by May next year. FYI, traditional 510k takes 6 months for review after submission. d. For coverage not yet cleared, you can still try them out for non-clinical purpose. Other countries EU MDR (Class IIb) : audit in progress (Stage 1: Jan 2023, Stage 2: Feb 2023), Asia SwiftMR FAQ 6  Vietnam - done Indonesia - done Hong Kong - N/A. Voluntary registration in progress Singapore - in progress Malaysia - in progress Thailand - in progress Taiwan- in progress Japan - in progress China - in progress LATAM Brazil - in progress Chile - N/A. IQ & Performance Scan time reduction 1. How far can you go with scan time reduction? a. As of today, we are confident with 50% reduction in scan time to be clinically acceptable. b. Some customers prefer to start off a little slow at around 35% reduction. As they build confidence in the product, they would push it up to the highest limit like 49%. SwiftMR FAQ 7  c. We have participated in global fastMRI challenges hosted by Facebook and NYU Langone in 2019 and 2020 where acceleration was tried up to 4 times and 8 times. Although we topped the challenge for two consecutive years, the outputs were not at a clinically acceptable level. That is, regardless of technical feasibility we take into consideration the clinical viability when turning it into a product. Accuracy & Reliability 1. How accurate is it? a. You would be familiar with sensitivity and specificity for CAD (Computer Aided Detection) softwares which tell you whether this person has a certain disease or not at what possibility. However, for SwiftMR, its function is not to detect but improve image quality. b. In terms of image quality, indices like SNR (Signal to Noise ratio), CNR (Contrast to Noise ratio), SSIM (Structural Similarity Index Measure), RMSE (Root mean square error) could be used for quantitative analysis. c. SSIM stands for Structural Similarity Index Measure. When we have both the original versus accelerated and enhanced images to compare, structural similarity between the two would be 1 if identical. SwiftMR’s SSIM scores surpass 0.97 at ease. But bear in mind that the original scan and accelerated scans are actually two different scans in time even if they are of the same person meaning that achieving 1 is inherently impossible. SwiftMR FAQ 8  2. Then how do you measure the performance? a. SwiftMR has been FDA 510k cleared for denoising and sharpness increase. SwiftMR increases the SNR of the input image by minimum 40%. For sharpness increase, FWHM (Full Width Half Maximum) is decreased by 0.43 to 4.5%. 3. How do you control the performance level? a. We take quality control very seriously. There are 2 steps to it. b. First is quantitative control where we carry out inhouse performance tests that measure SNR and sharpness increase levels for every deep learning model upgrades. c. Second is qualitative control where we carry out blind tests and/or explicit readings with external radiologists to verify whether image quality is acceptable on a regular basis. This is an example where we carried out a blind test with US board certified radiologists and asked them to compare blinded images based on all these specific questions. The blue and dark blue bars show that SwiftMR processed images were equivalent or even better than the original in more than 90% of the cases. SwiftMR FAQ 9  4. How reliable is the product? a. We carry out dozens of clinical researches with renowned partners to validate SwiftMR’s clinical efficacy in various conditions especially with abnormal cases. The results have been very promising so far. SwiftMR FAQ 10  b. Please refer to these case studies for Brain and Spine. Case report | Brain Case report | Spine c. But as with any AI SW, the most powerful evidence of reliability is the real world use case. Currently SwiftMR is used in more than 120 sites in Korea as of today. More than 50,000 exams are processed per month and accumulated total has exceeded 300,000 exams accumulated with zero down-time and zero claims. Missing information 1. If you shorten the scan time, are you not missing anything? a. If you are familiar with MRI, you would know that MRI is not at all like an X-ray or Ultrasound. The raw data of an MRI, kspace, is a collection of signals made up of frequencies and phases. Collecting a line less in kspace does not necessarily mean that I missed a certain part of the anatomy. Let’s put it that each location in kspace holds different attributes to an image. We collect the mission critical ones and let go on the ones that could be made up for later. For example, we take all the information regarding the structure of the image, but less on signal intensity and/or resolution. SwiftMR FAQ 11  2. But everyone’s inside looks different. How does the DL engine know what is anatomy (=signal) or noise? a. Well, that’s the secret source of our award-winning, proprietary DL engine. But we also understand that you need to know at least how it works to feel comfortable with it.3 b. SwiftMR’s main function is to recognize noise and remove it. Let’s take a Noise2Noise framework as an example. When the noise of inputs and labels has independent distributions with zero-mean, the network can be trained to have a similar performance to the network trained with clean labels using noisy input-label pairs only. Expanding this framework, we can leverage the noise independence in MR k-space data to generate training input-label pairs of two noisy independent k-spaces. Fast forward a thousand times, then you could get a clean image. c. We do not use GAN. 3. How can you be sure that the output image is not missing anything such as small lesions? a. We were also curious and eager to find out. A recent paper published in European Radiology is a fine example of how robust our deep learning engine could be with even submillimeter resolution. b. From the study, we were able to conclude that our deep learning framework successfully improved conventional 3D high-resolution MRI in all image quality parameters, fine structure delineation, and lesion conspicuity. Deep neural network application to 3D MRI whose pulse sequences and parameters were different from the training data showed improvement in image quality, revealing the potential to generalize on various clinical MRI. c. To utilize the spatial information of anatomical brain structures, DNN architecture is designed in a 2.5D fashion (slice concatenation along the channel dimension for inputs). Seven slices were concatenated for the network inputs, and the corresponding outputs were targeted to reconstruct the center slice of the inputs. d. For further details, please refer to the full paper. SwiftMR FAQ 12  [European Radiology] MR-self Noise2Noise: self-supervised deep learning– based image quality improvement of submillimeter resolution 3D MR images Artifacts 1. Can SwiftMR reduce artifacts? a. It can help reduce ringing (=truncation) artifacts. b. It cannot reduce other artifacts that already exist in the original image. However, by reducing the actual scan time, it can help reduce the possibility of motion artifact occurring in the first place. Also, we can help optimize the scan parameters so that the artifact occurs less in the original image. 2. Can SwiftMR introduce new artifacts? a. No, it does not introduce new artifacts. However, in the process of increasing sharpness for normal structures, it may give distinction to an artifact that already exists in the original image. In this case, we can help you adjust the sharpness level or the original scan parameter to minimize the artifact from the MRI scanner. Preferences 1. The image looks cartoonish, too smooth, too clean, somewhat artificial. a. This is usually when the image has been denoised far more than what you are used to. b. We offer different denoising levels and can adjust it down to where you are comfortable. 2. What if I don’t like the image? a. We know that all eyes are different which is why we offer all new customers with a one-month free trial. Try it for yourself then decide. A dedicated Customer Success Manager will be allocated to support you with image optimization. b. We are very open to feedback. If you have specific feedback on the product quality as you use the product, please feel free to provide them to us. We will review them and if found to be a valid request, we will reflect it into the next earliest release. SwiftMR FAQ 13  c. Other than that, if the image quality issue is found to be originating from the MR scanner itself, we won’t be able to support you. Please contact your MR vendor. d. For whatever other reason, if you don’t like it, we recommend that you don’t use it. Deep Learning 1. How is it trained? a. We train the model with pairs of low quality image and high quality image. After substantial training, the model acquires the ability to infer from a low quality image the most adequate high quality image. 2. Do you use supervised or unsupervised learning? a. Let’s say a bit of both. 3. How big is the training data set? a. Around 3 million images. In terms of patients, 80,000 patients. 4. Where do you get the data from? a. Multiple sources including IRB-approved clinical research studies, DRB- approved collaboration partners, paid platforms, and open platforms. 5. Will it learn from my data when deployed on my site? a. No, SwiftMR does NOT learn live in real world. b. All the training happens under our roof with images from compliant data sources. Only when the model passes the performance test, it will be frozen and released on our server. Workflow & Installation Workflow 1. How does it work? SwiftMR FAQ 14  1. SwiftMR sits in between the MRI and PACS. That is, nothing is installed on the MRI or on PACS. They are just connected through network using IP addresses. 2. Without SwiftMR, MRI would send the DICOM images to PACS. With SwiftMR, MRI would instead send to SwiftMR, SwiftMR would enhance the image, then send to PACS. 2. As an MR tech, will my workflow change? a. No, the workflow remains the same for MR techs and radiologists. Scan the patient, check the image on the MRI PC, then export to SwiftMR. The rest will be automatically done. b. When you check the image after the scan, it may be a little noisy, but you can still tell if rescan is needed or not due to motion. 3. How does SwiftMR know which images to process? a. SwiftMR will only process the series names that it recognizes. b. Upon installation, our service engineer would register the series names that you have chosen to use for your site. c. For example, if the series description includes a prefix or suffix “Swift”, the image is processed. If the name doesn’t match, it will just be passed through without any processing. If a study is comprised of both supported and unsupported series, you’ll get the whole study with only the ones supported processed. d. For the naming convention, you can use any keyword you like. It just needs to be mapped and confirmed with our sever. SwiftMR FAQ 15  4. What will be changed on my protocol? a. We don’t change anything on your original protocol. b. You’ve got two choices. If you are currently unhappy with the image quality, then just register your current protocol to be processed with SwiftMR. Then you get enhanced image quality with same scan time. c. However, if you’d like to see reduction in scan time, we can support you create a new set of protocols. This will be a duplicate of your original protocol where only the parameters related to scan time will be changed such as Average, Grappa (Parallel imaging factor), Bandwidth, Echo train length, oversampling. Optimization will be supported until image quality meets your expectations. 5. What if I want to keep both the before/after images? a. We offer the option of saving both the before/after SwiftMR images or just after images. b. We observe that customers tend to save them both during the free trial to compare them side by side. Then turn this function off as they gain confidence in the product and only keep the final version. c. If you choose to send only the final version to PACS, the before image would still be available on your MRI for a certain period of time. 6. How long does the whole enhancing process take? a. A couple of seconds for a short sequence and up to 1 minute for a whole MR study to appear on your PACS viewer fully processed. For the worst case of up to 300 images in a study (10 sequences * 30 slices) can take up to 2 minutes. Total time may vary slightly depending on the network speed of the institution. b. A little more time required for 3D images. Installation 1. How long does it take to install? a. Actual setup takes less than 3 hours, but we usually plan 2 full days for implementation. SwiftMR FAQ 16  b. Day 1 is for network configuration and MRI protocol set up. Less than an hour for network configuration and around 2 hours per MRI for MRI protocol setup depending on the number of sequences. Our customer success manager would need full access to your MRI console for during the set up. c. On Day 2, you can start using the product after user training and our CSM would be monitoring onsite. If any issues are found during the operation such as connectivity or image quality, CSM can support you immediately. 2. Do I need to install any hardware? a. If you are open to using cloud service, then no. All computing power required to run the deep learning engine will be provided through cloud. We do provide a mini pc that acts like a gateway for secure network communication. This is complimentary with the subscription. b. If you are NOT open to cloud, then on-premise server deployment is also available. This would require a decent sized workstation to be installed on site to carry out the deep learning inference offline. 3. Would it be possible to send back to MRI scanner? a. Yes. Our implementation method is very flexible. As long as our destination supports DICOM communication and you can provide us IP, Port, and AE Title, we can communicate with any workstation. So MRI → SwiftMR → MRI is possible. b. Also, SwiftMR supports multi-destination. For example, after processing, enhanced image can be sent to both MRI scanner and PACS server. Cloud SwiftMR FAQ 17  1. Which CSP (Cloud Service Provider) do you use? a. We are on AWS (Amazon Web Services). 2. Where is the data region? Where will my data be handled? a. We utilize multiple locations to serve our global customer base. b. For US, we have a separate instance on US East, North Virginia. 3. Is it HIPAA compliant? a. Yes, the product as well as our company is HIPAA compliant. b. Upon adoption, we will be asking you to sign a BAA (Business Associate Agreement) with us which allows us to handle your institution’s PHI(Protected Health Information) for the specific purpose of providing image enhancement service for MR images. 4. Is it safe with cloud? a. We have designed it as safe as possible by removing all sensitive PHI(Protected Health Information) from the file when sending off to cloud for image enhancement. There is a function in the gateway pc where it splits off PHI from the metadata of the DICOM file, labels it with our own identification number, and sends only the pixel image to cloud for processing. When it returns, it will match it up with the PHI left behind, save the complete DICOM file in PACS. b. Encryption is default for all files in transit and at rest. Within the hospital network the system follows DICOM protocol and over the internet HTTPS/TSL/SSL protocol. c. We do not save any files on our server. The files on our cloud server are permanently deleted after 24 hours. d. Our cloud server is equipped with IPS (Intrusion protection system), WAF (Web Application Firewall), 24x7 monitoring. Competitive landscape 1. Are there any competitors that provide similar products? SwiftMR FAQ 18  a. Yes, but not many. One group is of MR vendors like GE, Siemens, and Philips providing AI SWs on their scanners, and the other group is 3rd party vendors like us. We are from Korea out of Seoul National University, Subtle Medical is US-based, out of Stanford, and Medic Vision is out of Israel and active in the US. 2. How do you compete against MR vendors? a. We acknowledge that what the vendor provides would have better usability since the software would be embedded into the scanner. However, it would only be provided to the latest models, or require an expensive upgrade to the hardware for a bit older models. Even when you do upgrade, it will be just for that scanner. b. With our solution, you can virtually upgrade the whole fleet of MRIs regardless of vendor or model or age without having to upgrade hardware for each one of them. We have regular upgrades to our model providing the latest performance enhancement to you at no extra cost. Think of the last time you got your software upgrade for your MR scanner. And think of how fast things move with AI. We won’t leave you out in the cold. 3. How do you compare against other 3rd party SW vendors? a. Better in all aspects of the product: image quality and usability. b. Monitoring UI provided so that you can monitor the progress. c. Other than that, we believe customer experience also matters a lot. So we’ve just started building our local team here based in Chicagoland and soon enough we’ll be ready to offer our product and service to all the states. Customer support General 1. If something goes wrong, who do I contact? a. You’ll have a dedicated Customer Success Manager (CSM) to open cases. 2. How long would it take to resolve the issue? SwiftMR FAQ 19  a. We are pretty quick with turnaround time. If you are open to remote access, it will be the quickest way for us to look into your issues. Less than an hour on average and up to 4 hours for complex issues. b. For other cases that require on-site visits, we’ll try to arrange a schedule at the earliest. Liabilities 1. Does SwiftMR impact the MR scanner in any way? a. No, SwiftMR is a stand-alone solution that does not impact any other device or software. Communication with other devices such as MRI or PACS follows the usual DICOM protocol, so it is recognized as another healthcare modality by other devices. b. Regarding the parameter configurations on the MRI console, it is carried out within the range allowed to the user by the vendor. By reducing the scan time, it has the effect of extending the lifespan of the machine as you are running it at a less severe condition. c. We have not had any case where SwiftMR adversely impacts MR scanners across all of our install bases which reaches more than 130 sites. 2. Do you provide warranties on issues arising from image quality, patient claims, cybersecurity? a. Yes, if it can be proven that SwiftMR is the cause of the issue, we will be fully liable. b. SwiftMR is covered by Product Liability insurance for medical claims and Premier Tech insurance for cybersecurity claims by CHUBB, the world’s biggest insurer, in all the global markets that we enter. SwiftMR FAQ 20