The future of MRI in radiation therapy: Challenges and opportunities for the MR community

Magnetic Resonance in Medicine, Год журнала: 2022, Номер 88(6), С. 2592 - 2608

Опубликована: Сен. 21, 2022

Abstract Radiation therapy is a major component of cancer treatment pathways worldwide. The main aim this to achieve tumor control through the delivery ionizing radiation while preserving healthy tissues for minimal toxicity. Because relies on accurate localization target and surrounding tissues, imaging plays crucial role throughout chain. In planning phase, radiological images are essential defining volumes organs‐at‐risk, as well providing elemental composition (e.g., electron density) information dose calculations. At treatment, onboard informs patient setup could be used guide placement sites affected by motion. Imaging also an important tool response assessment plan adaptation. MRI, with its excellent soft tissue contrast capacity probe functional properties, holds great untapped potential transforming paradigms in therapy. MR Therapy ISMRM Study Group was established provide forum within community discuss unmet needs fuel opportunities further advancement MRI applications. During summer 2021, study group organized first virtual workshop, attended diverse international clinicians, scientists, clinical physicists, explore our predictions future next 25 years. This article reviews findings from event considers challenges reaching vision expanding field.

Язык: Английский

---

U-net architecture with embedded Inception-ResNet-v2 image encoding modules for automatic segmentation of organs-at-risk in head and neck cancer radiation therapy based on computed tomography scans

Physics in Medicine and Biology, Год журнала: 2022, Номер 67(11), С. 115007 - 115007

Опубликована: Фев. 8, 2022

Purpose.The purpose of this study was to utilize a deep learning model with an advanced inception module automatically contour critical organs on the computed tomography (CT) scans head and neck cancer patients who underwent radiation therapy treatment interpret clinical suitability results through activation mapping.Materials methods.This included 25 that were delineated by expert oncologists. Contoured medical images 964 sourced from publicly available TCIA database. The proportion training, validation, testing samples for development 65%, 25%, 10% respectively. CT segmentation masks augmented shift, scale, rotate transformations. Additionally, pre-processed using contrast limited adaptive histogram equalization enhance soft tissue while contours subjected morphological operations ensure their structural integrity. based U-Net architecture embedded Inception-ResNet-v2 blocks trained over 100 epochs batch size 32 rate optimizer. loss function combined Jaccard Index binary cross entropy. performance evaluated Dice Score, Index, Hausdorff Distances. interpretability analyzed guided gradient-weighted class mapping.Results.The mean Distance averaged all structures 0.82 ± 0.10, 0.71 1.51 1.17 mm respectively data sets. Scores 86.4% compared within range or better than published interobserver variability derived multi-institutional studies. average training time 8 h per anatomical structure. full anatomy network required only 6.8 s patient.Conclusions.High accuracy obtained large, set, short clinically-realistic prediction reasoning make proposed in work feasible solution scan environment.

Язык: Английский

---

Uncertainty Assessment for Deep Learning Radiotherapy Applications

Seminars in Radiation Oncology, Год журнала: 2022, Номер 32(4), С. 304 - 318

Опубликована: Окт. 1, 2022

In the last 5 years, deep learning applications for radiotherapy have undergone great development. An advantage of over radiological is that data in are well structured, standardized, and annotated. Furthermore, there much to be gained automating current laborious workflows radiotherapy. After initial peak belief learning, researchers also identified fundamental weaknesses learning. The basic assumption training test originate from same generating process. This not always clear-cut clinical practice, eg, acquired with 2 different scanners vendors might it important realize residual uncertainties remain even if arise process as data. As being introduced workflows, a model must express user when prediction exceeds certain uncertainty threshold. literature on assessment still its infancy; however, quite body exists validity models computer vision applications. paper tries explain these general concepts community. Concepts epistemic aleatoric techniques them described detail. It discussed how they can applied maximize confidence automated learning-driven workflows. Their usage demonstrated 3 examples applications, ie, dose prediction, synthetic CT generation, contouring. final part, some key elements ensure automatic alerting missing discussed. State-of-the-art solutions checking within-distribution vs out-of-distribution samples However, methodologies immature, strict QA protocols close human supervision will needed. Nevertheless, offer already value

Язык: Английский

---

Benefits of automated gross tumor volume segmentation in head and neck cancer using multi-modality information

Radiotherapy and Oncology, Год журнала: 2023, Номер 182, С. 109574 - 109574

Опубликована: Фев. 22, 2023

PurposeGross tumor volume (GTV) delineation for head and neck cancer (HNC) radiation therapy planning is time consuming prone to interobserver variability (IOV). The aim of this study was (1) develop an automated GTV approach primary (GTVp) pathologic lymph nodes (GTVn) based on a 3D convolutional neural network (CNN) exploiting multi-modality imaging input as required in clinical practice, (2) validate its accuracy, efficiency IOV compared manual setting.MethodsTwo datasets were retrospectively collected from 150 cases. CNNs trained with consensus ground truth, either single (CT) or co-registered multi-modal (CT + PET CT MRI) data input. For validation, GTVs delineated 20 new cases by two observers, once manually, correcting the delineations generated CNN.ResultsBoth performed better than single-modality CNN selected validation. Mean Dice Similarity Coefficient (DSC) (GTVp, GTVn) respectively between (69%, 79%) (59%,71%) MRI. DSC corrected (81%,89%) (69%,77%) observers (76%,86%) (95%,96%) delineations, indicating significant decrease (p < 10−5), while increased significantly (48%, p 10−5).ConclusionMulti-modality HNC shown be more efficient consistent setting beneficial over approach.

Язык: Английский

---

Dosimetric comparison of autocontouring techniques for online adaptive proton therapy

Physics in Medicine and Biology, Год журнала: 2023, Номер 68(17), С. 175006 - 175006

Опубликована: Июнь 29, 2023

Abstract Objective. Anatomical and daily set-up uncertainties impede high precision delivery of proton therapy. With online adaptation, the plan is reoptimized on an image taken shortly before treatment, reducing these and, hence, allowing a more accurate delivery. This reoptimization requires target organs-at-risk (OAR) contours image, which need to be delineated automatically since manual contouring too slow. Whereas multiple methods for autocontouring exist, none them are fully accurate, affects dose. work aims quantify magnitude this dosimetric effect four techniques. Approach. Plans automatic compared with plans contours. The include rigid deformable registration (DIR), deep-learning based segmentation patient-specific segmentation. Main results. It was found that independently method, influence using OAR small (<5% prescribed dose in most cases), DIR yielding best Contrarily, contour larger (>5% indicating verification remains necessary. However, when non-adaptive therapy, differences caused by were coverage improved, especially DIR. Significance. results show adjustment OARs rarely necessary several techniques directly usable. important. allows prioritizing tasks during time-critical adaptive therapy therefore supports its further clinical implementation.

Язык: Английский

---