Whitepaper AI-Rad Companion Organs RT - The power of automated contouring at CCGM Montpellier

Customer Experience AI-Rad Companion Organs RT The power of automated contouring at CCGM Montpellier siemens-healthineers.com/ai-rad-companion Rendering not generated in the AI-Rad Companion Organs RT SIEMENS Healthineers Whitepaper · AI-Rad Companion Organs RT Table of contents Challenges with Organs at Risk contouring …………………………………………… 3 AI-Rad Companinon Organs RT – Overview …………………………………………… 4 Automatic contouring for cancer therapy ……………………………………………… 5 Automatic contouring of OAR: a study with CCGM Montpellier ………………… 6 Appendix …………………………………………………………………………………………… 9 Understanding the mechanism behind automatic contouring ……………… 10 References 11 ……………………………………………………………………………………… 2 September 2021 · siemens-healthineers.com/ai-rad-companion AI-Rad Companion Organs RT · Whitepaper Challenges with Organs at Risk (OAR) contouring In many institutions, OAR are contoured contouring still posses a challenge for manually and as a result, valuable staff many institutions. resources are tied up, turning OAR contouring into a cost and time intensive task. In addition, In the last decade, various automatic contour- inter-observer variability can make it difficult ing solutions have been introduced to address to achieve consistent contouring results and these challenges. However, the results may operators need to be trained on common not be clinically useful for the RT professionals contouring guidelines. Considering staffing leading to significant editing or redoing of issues such as high turnover rates, OAR the contours. Inconsistent No interobserver variability up to 1 hour/patient [1] Time consuming Potential for for OAR contouring time savings Figure 1: Manual contouring is a time consuming process requiring considerable effort which often leads to significant variability and inconsistency between users during the radiotherapy planning process. September 2021 · siemens-healthineers.com/ai-rad-companion 3 Whitepaper · AI-Rad Companion Organs RT AI-Rad Companion Organs RT – Overview Automatic contouring for AI-Rad Companion Organs RT What is AI-Rad Companion Organs RT? head and neck, thorax, abdomen, and pelvis. AI-Rad Companion Organs RT is a an AI-based It also supports the use of organ template solution that provides radiation oncology configurations that can be aligned with institu- professionals with automatic contouring of tional protocols; this may save time and organs at risk, which is input to their radiation improve standardization in clinical workflows. therapy planning via the teamplay digital health platform. The images acquired at the CT What are the key benefits? scanner are sent to AI-Rad Companion Organs By leveraging Artificial Intelligence (AI) to RT to be processed, and then the RTStruct generate OAR contouring, AI-Rad Companion (DICOM) results can be pushed directly to the Organs RT, enables high quality OAR contour- treatment planning system or first reviewed ing to drive standardization with AI-powered in the AI-Rad Companion Organs RT interface. algorithms. These benefits can potentially free up staff to spend more time on other AI-Rad Companion Organs RT provides organs tasks and help to simplify radiotherapy plan- at risk contours using deep-learning (AI) ning workflow. algorithms for various body regions, including Al-Rad Companion hosted on teamplay digital health platform Treatment Planning System Local T Cloud Organs-at-risk contouring > Efficiency Abdomen Brain, Head & Neck > Consistency Prostate > Accessibility Breast Lung Figure 2: AI-Rad Companion Organs RT, seamlessly integrates into your hospital environment – and your workflows. AI-Rad Companion Organs RT is not commercially available in all countries. Future availability cannot be ensured. 4 September 2021 · siemens-healthineers.com/ai-rad-companion AI-Rad Companion Organs RT · Whitepaper Al-Rad Companion Organs RT: Organs-at-risk autocontouring trained by deep Learning Head & Neck Breast Lung Abdomen Prostate Multiple other organs supported . Spinal cord Esophagus . Skeleton . Body contours Al-Rad Companion Organs RT is not commercially available in all countries and its future availability cannot be ensured. Figure 3: Overview of body parts and cancer sites that are supported by AI-Rad Companion Organs RT. HeQELEMENS A-Rod Companion Organs RT Organe AT contoured the folowing organs based on Template : Scan: 2019-11-21 11:40 Esophagus Female Breast Left Female Breast Right DICOM -RT Structure Set Lung Late night Upper Decline sand to TPS Show All Contours Figure 4: The result preview displays a contoured dataset of the thorax. Figure 5: Exemplary visualization of a contoured region in the male pelvis. Courtesy: University Hospital Erlangen, Erlangen, Germany Courtesy: University Hospital Erlangen, Erlangen, Germany September 2021 · siemens-healthineers.com/ai-rad-companion 5 Whitepaper · AI-Rad Companion Organs RT Automatic contouring for cancer therapy In the last couple of years, not only has the the increase in the number of patients puts cancer incidence rates increased, but so has significant pressure on radiotherapy staff the amount of patients receiving Radiation responsible for OAR contouring results. In Therapy (RT). Up to two thirds of all patients addition, effective treatment planning for with cancer will need RT treatment during cancer patients, requires high quality contours. the course of their disease [2]. Advances in technology and AI can potentially help automate repetitive tasks such as OAR Each patient arriving at the RT department contouring to reduce workload, save time requires a treatment plan. Contouring the and drive standardization using AI-powered organs-at-risk is the necessary first step in algorithms. the process of treatment planning. Therefore, Increase in worldwide cancer Percentage of cancer patients receiving cases expected radiation therapy Total of new cases worldwide [3] Up to 2/3 of cancer patients receive RT [4] +33.4% 66% 24.1 million 18.1 million 2018 . Cancer patients . 2030 Cancer patients receive RT Figure 6: Cancer statistics and incidence predictions. 6 September 2021 · siemens-healthineers.com/ai-rad-companion AI-Rad Companion Organs RT · Whitepaper Automatic contouring with AI-Rad Companion Organs RT: a study with CCGM Montpellier Overview Improving therapeutic decision and treat- “Automatic contouring helps with ment planning in radiation oncology increase precision which is highly The manual delineation of organs at risk (OAR), is a time-consuming labor-intensive process beneficial for patients.” * prone to inter-observer variability. In addition, the ability to adapt radiation treatments during the therapy phase limits the optimal use of Process & methods adaptive radiotherapy. Automating the con- The automated contours generated by AI-Rad touring process, can be beneficial in optimizing Companion Organs RT were evaluated by two workflows and enabling efficient radiation experienced board certified clinicians (a physi- therapy planning in a patient care pathway. cian and physicist). An agreed upon study was Effective treatment planning requires high defined where clinicians would analyze con- quality contour results, to ensure a more tours generated by AI-Rad Companion Organs precise dose distribution and optimization RT of the Head & Neck, Thorax (Breast/Lung), of radiotherapy treatments thereby sparing Abdomen, Pelvis (Male/Female) and compare healthy OAR located near a tumor, from to contours generated by skilled personnel, unnecessary radiation. according to CCGM’s current standard clinical processes and provide feedback. A 4-point AI-powered automatic contouring scale was used to judge the AI-Rad Companion To analyze the AI-powered automatic contour- Organs RT generated contours, where a rating ing cloud based service of the AI-Rad of 4 is clinically usable, 3 – contours required Companion Organs RT, a clinical evaluation minor edits, 2 – contours required major edits, was conducted together with Oncology and a rating of 1 is need to re-do. This scale institution, Centre de Cancérologie du Grand was used to simplify contour grading. To deter- Montpellier (CCGM) located in the southern mine time-savings, manual contouring times region of France. The entire study was were assessed in a sample set of cases where conducted over a five month period. the involved clinicians measured the time it took to manually contour the OARs. The aim of the clinical evaluation was to con- firm that the results of automatic contouring “The automatic contouring feature were in line with the expectation from CCGM on the functionality of AI-Rad Companion of the AI-Rad Companion Organ RT Organs RT across three key areas: is life changing for us clinicians.” * • Time-savings using through automated organs at risk contouring Manual contouring tools such as interpolation • Standardization with AI-powered algorithms were allowed. These times where then com- for high-quality organs at risk pared to the time spent in the automated OAR • Simplifying clinical workflow by automati- contouring to estimate time-savings. cally providing OAR contours to the Treatment Planning System (TPS) * All quotes in this document are from Prof. Muraro. The statements by Siemens Healthineers’ customers described herein are based on results that were achieved in the customer’s unique setting. Because there is no “typical” hospital or laboratory and many variables exist (e.g., hospital size, samples mix, case mix, level of IT and/or automation adoption) there can be no guarantee that other customers will achieve the same results. September 2021 · siemens-healthineers.com/ai-rad-companion 7 Whitepaper · AI-Rad Companion Organs RT Results & discussion Summary Of the OAR contours for combined regions AI-Rad Companion Organs RT is designed to generated by AI-Rad Companion Organs RT address the labor-intensive and time consum- and evaluated by CCGM (N = 55), as shown ing segmentation of OARs which requires high in figure 7, 77% of generated contours quality results for optimal radiation therapy required no edits and were clinically usable planning. Leveraging an AI based deep learn- (rating = 4) and 95% of the generated ing algorithm deployed in the cloud, AI-Rad contours were clinically usable or required Companion Organs RT is able to seamlessly minor edits (ratings 4 & 3). perform automatic contouring for output to a treatment planning system in the form The Thorax, male and female pelvis, received of DICOM RT structure set files (RTStruct). rating of greater than 95% results which include contours that are clinically usable Overall the results from the study with CCGM, or require minor edits. showed acceptable outcomes in the automatic contouring of OARs using AI-Rad Companion Across all organs covered in this study, only Organs RT. The annotators evaluation showed 7–8% generated contours required a repeat that close to 77% of organ contours required with a must redo rating (rating = 1). no modification and 95% of the organs contoured were classified as either clinically usable or requiring minor edits. 100% The AI-Rad Companion Organs RT solution, 90% supports a fully automated workflow with the potential for time-savings and increased 80% standardization with the use of AI-powered algorithms for high quality organs at risk 70% contouring. Time saved with automated OAR 60% contouring may allow clinicians to be freed up to focus on other clinical tasks. 50% 40% 30% > 95% 20% Clinically usable or minor edits 10% 0% All Head Thorax Abdomen Male Female and Neck pelvis pelvis 1 must redo 2 major edits Head & Thorax Abdomen Pelvis 3 minor edits Neck (breast & lung) 4 clinically usable Figure 7: Contour ratings. Figure 8: Organs in review. 8 September 2021 · siemens-healthineers.com/ai-rad-companion AI-Rad Companion Organs RT · Whitepaper Appendix Training of AI-Rad Companion Organs RT algorithm The AI-Rad Companion Organs RT algorithm training, CT datasets were obtained for each is trained to leverage Deep Learning (DL) body region from various radiation therapy and technology that enables supervised learning. radiology departments in Europe and America. By exploiting supervised and unsupervised deep learning methods, AI Rad Companion Ground-truth contours were manually Organs RT algorithm is trained with high generated on these CT datasets by a team of quality data annotated by experts to achieve experienced annotators overseen by radiation robust performance. oncologists or radiologists. For this process, a consistent annotation protocol was set In order to learn organ segmentation, a Deep up beforehand based on widely accepted Image-to-Image Network (DI2IN) is employed. consensus guidelines such as the ones pub- It consists of a convolutional encoder-decoder lished by the Radiation Therapy Oncology architecture combined with a multi-level feature Group (RTOG/NGR). The deep learning model concatenation. A Generative Adversarial for organs segmentation is trained with Network (GAN) is selectively used to regularize optimized CT data and corresponding ground the training process of DI2IN by discriminating truth segmentation following standardized the prediction of the DI2IN from the ground annotation protocol. truth. The model is selected in the epoch with the best performance on the validation AI-Rad Companion Organs RT employs an set. A GAN uses two networks that compete automatic contouring process where deep against each other during the training process. learning AI algorithms are used to provide The first network – the generator – tries to organs at risk (OAR) contouring on CT images emulate a human drawn contour while the according to pre-selected structure templates. second network – the discriminator – tries to Structures with pre-defined attributes are discriminate the prediction of the first network stored within the organ database. The tem- from the ground truth (human drawn contour). plate configuration feature, defines the organs The information is then fed back to the to be contoured and sets rules for the automat- respective networks. This iterative process ic selection of templates. The results of the ensures that during the training of the automated contouring generated by AI-Rad networks, the machine generated contours Companion Organs RT, can be sent directly become virtually indistinguishable from to a treatment planning system (TPS) thereby the human drawn contours. For algorithm enabling a seamless workflow. September 2021 · siemens-healthineers.com/ai-rad-companion 9 Whitepaper · AI-Rad Companion Organs RT Understanding the mechanism behind automatic contouring of OARs The automatic process of contouring network (DRL) [5]. The result is a cropped im- OARs relies on a deep learning based age of the target organ region. In the second model which consists of a two step step, the cropped image is used as input to approach as seen in Figure 9. create the contours. This step is based on a DI2IN [6]. The DI2IN was trained to its optimal performance in the Siemens Healthineers In the first step, the target organ region in AI environment. The DRL algorithm also has the optimal input image is extracted using the ability to detect multiple target regions. a trained Deep Reinforcement Learning Step 1: Locate target organ region Input image Deep Reinforcement Learning [5] Cropped image Step 2: Contour target organ Cropped image DI2IN [6] Segmented organ Figure 9: Two step algorithm for DL-based contouring. Please note – the algorithm is not self-learning. Your data is not used for further training. 10 September 2021 · siemens-healthineers.com/ai-rad-companion AI-Rad Companion Organs RT · Whitepaper References [1] Nambu A, et al. Rib fracture after stereotactic radiotherapy for primary lung cancer: prevalence, degree of clinical symptoms, and risk factors. BMC cancer, 2013, 13. Jg., Nr. 1, S. 68. [2] American Cancer Society, www.cancer.org [3] International Agency for Research on Cancer, https://gco.iarc.fr/tomorrow/home, accessed July, 2020 [4] RTAnswers.com, https://www.rtanswers.org/ What-is-Radiation-Therapy, accessed July, 2020. [5] Ghesu FC, et al. Multi-scale deep reinforcement learning for real-time 3D-landmark detection in CT scans. IEEE transactions on pattern analysis and machine intelligence, 2017, 41. Jg., Nr. 1, S. 176-189. [6] Yang D, et al. Automatic Liver Segmentation Using Adversarial Image-to-Image Network. U.S. Patent Application Nr. 15/877,805, 2018. September 2021 · siemens-healthineers.com/ai-rad-companion 11 Siemens Healthineers Headquarters Siemens Healthcare GmbH Henkestr. 127 91052 Erlangen, Germany Phone: +49 9131 84-0 siemens-healthineers.com Published by Siemens Healthcare GmbH · Online · ©Siemens Healthcare GmbH, 2021