Using Deep Learning for MRI to Identify Responders to Chemoradiotherapy in Rectal Cancer
2020; Radiological Society of North America; Volume: 296; Issue: 1 Linguagem: Inglês
10.1148/radiol.2020200417
ISSN1527-1315
Autores Tópico(s)Colorectal Cancer Screening and Detection
ResumoHomeRadiologyVol. 296, No. 1 PreviousNext Reviews and CommentaryFree AccessEditorialUsing Deep Learning for MRI to Identify Responders to Chemoradiotherapy in Rectal CancerDow-Mu Koh Dow-Mu Koh Author AffiliationsFrom the Department of Radiology, Royal Marsden Hospital, Downs Road, Sutton SM2 5PT, England.Address correspondence to the author (e-mail: [email protected]).Dow-Mu Koh Published Online:Apr 21 2020https://doi.org/10.1148/radiol.2020200417MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Zhang and Wang et al in this issue.Dr Dow-Mu Koh is professor of functional cancer imaging at the Royal Marsden Hospital and Institute of Cancer Research. His research interests include body diffusion-weighted imaging and abdominal malignancies. He was associate editor of Radiology for MRI from 2011 to 2018. Dr Koh is a fellow of the ISMRM, ESGAR, and ICIS. He is director of the NIHR Clinical Research Facility at the Royal Marsden Hospital.Download as PowerPointOpen in Image Viewer Rectal cancer is common, and surgery is potentially curative. However, the surgical outcome is dependent on whether the tumor can be completely removed at surgery. In patients with locally advanced disease, there is high risk of leaving tumor tissue behind after surgery. For this reason, neoadjuvant chemoradiotherapy (NCRT) is usually administered to downsize and downstage these tumors prior to surgery to improve the chance of complete surgical clearance.Interestingly, between 10% and 25% of such cancers might be cured by NCRT alone (ie, cancers with pathologic complete response [pCR]). Unfortunately, confident identification of pCR at imaging after neoadjuvant treatment is challenging. This is because residual or treatment-related signal changes on MRI often result in disease overstaging (1). This means that patients with pCR still might be referred for bowel surgery, with its associated morbidity, mortality, and functional impairment, without clinical benefit. For this reason, the ability to confidently identify pCR to NCRT in patients with rectal cancer is highly desirable.Conventional imaging assessment of response to NCRT relies on MRI. Thin-section small field-of-view T2-weighted MRI is used in clinical practice to stage primary rectal tumors (using TNM criteria), assess surgical resectability (by evaluating the relationship of the tumor to the potential circumferential resection margin), and identify poor prognostic imaging features (eg, extramural venous invasion and depth of extramural disease extension). After chemoradiotherapy, T2-weighted MRI is applied to assess treatment response by restaging the tumor (using TNM criteria) and to determine the tumor regression grade (TRG). The TRG (TRG0, TRG1, TRG2, and TRG3) is assessed by the extent to which intermediate tumor T2 signal intensity is replaced by low T2 signal intensity fibrosis, with TRG0 showing no residual tumor signal and TRG3 showing predominant tumor signal intensity. TRG has been shown to be a good predictor of pathologic response and disease survival (2–4). However, a recent meta-analysis (5) found that while TRG has good specificity (93%) for pCR, it has limited sensitivity (32%).Diffusion-weighted MRI has also been applied to evaluate disease response in rectal cancer. Diffusion-weighted MRI is used to evaluate the mobility of water protons in tissues. A good response to treatment is usually observed as a decrease in the tumor signal on diffusion-weighted MRI, accompanied by an increase in the tumor apparent diffusion coefficient (ADC). However, spurious high signal intensity may be observed in up to about 50% of patients at the site of proven pCR (1). In addition, the apparent diffusion coefficient is also unreliable in identifying patients with pCR to treatment (6). This is because of overlap in apparent diffusion coefficients between patients with pCR and those with noncomplete pathologic response.For these reasons, there has been interest in applying more complex diffusion models to improve disease response characterization. One model, known as diffusion kurtosis imaging (DKI), has been used to probe the non-Gaussian behavior of water diffusion in rectal cancer, which in theory can better reflect changes in microstructural organization and cellular integrity that occur with effective treatment. However, to apply DKI in the body requires imaging to be performed using at least three b values, with at least one b value being higher than 1500 sec/mm2 (7). This can translate to longer image acquisition time. Nonetheless, one small study using DKI found that the mean kurtosis value of rectal cancer measured after chemoradiotherapy showed high sensitivity (92%) and specificity (83%) in the identification of patients with pCR compared with conventional apparent diffusion coefficient measurements (8).Clearly, the ability to combine morphologic and functional imaging data to identify patients with pCR is of great interest. One way to achieve this is to make use of artificial intelligence to train a diagnostic model to undertake this task. With the increasing availability of graphic processing units and trained image networks, it is now feasible to apply deep learning to train a neural network to identify patients with pCR by using T2- and diffusion-weighted images as inputs. The performance of such an algorithm can then be tested in an independent validation study cohort and against the performance of radiologists performing standard image reading.In the study by Zhang et al (9), the authors recruited 383 patients with locally advanced rectal cancer from one institution. Participants underwent conventional T2-weighted MRI and DKI (using 12 b values) before and after NCRT. Images from the first 290 patients were used for the training cohort, while images from the remaining 93 patients constituted the validation cohort. All images were evaluated by two radiologists who drew regions of interest on images obtained with T2-weighted MRI and diffusion-weighted MRI with a b value of 1000 sec/mm2 to delineate the site and extent of tumors. A multipath convolutional neural network was used for model training, using eight types of image input comprising T2-weighted images, the diffusion coefficient map, the kurtosis map, and images with a b value of 1000 sec/mm2 obtained before and after neoadjuvant treatment. The models were trained to predict the class probabilities of (a) pCR versus noncomplete pathologic response, (b) low (TRG0 and TRG1) versus high (TRG2 and TRG3) TRG scores, and (c) T stage downgrading versus no T stage downgrading.Zhang et al found that the best-performing model of the three was the one used to discriminate between patients with pCR and those without (area under the receiver operating characteristic curve, 0.99; 95% confidence interval: 0.94, 1.00). This model outperformed the two radiologists in assessing the presence of pCR after treatment (rater 1, 73.1% accuracy; rater 2, 75.2% accuracy) based on standard image reading. However, the subsequent provision of the trained model outputs to the radiologists to support their image assessment significantly increased their diagnostic performance (rater 1, 87.1% accuracy; rater 2, 86.0% accuracy). The provision of the trained model also resulted in a reduction in the error rate (by 13% and 14%, respectively, for raters 1 and 2), false-positive prediction, and false-negative prediction of pCR in the study cohort. However, although the diagnostic performance for identifying patients with pCR was augmented by deep learning, it is worth noting that the diagnostic performance of the deep learning–trained models for classifying tumor regression grade and T stage downgrading was not better than that of the two radiologists at direct image reading.While the results from this study are highly encouraging, they also highlight a number of issues that warrant further consideration. First, DKI is not widely performed, and the very high–b-value images required for mathematical modeling are sensitive to poor signal-to-noise ratio. This is especially true since treated disease may return little signal on the high-b-value images. Hence, a clear understanding of the repeatability of the DKI-derived parameters in this disease setting would be helpful for its wider translation. Second, the deep learning model in this study has been trained using MRI images from a single institution. Given that there can be substantial signal intensity variations across images acquired from different studies and across different MRI platforms, it is unclear to what extent the current model can be widely generalized. Third, it is not known which features on the eight different images used to train the model are accounting for its high diagnostic performance. The ability to unravel the relative contributions of the different images to model performance outside a black-box environment could be helpful to gain insights that may aid future artificial intelligence developments.Future studies should consider using multicenter real-world data to retrain the model and assess its performance in a multicenter trial. If proven to be accurate, such an algorithm will no doubt have a substantial impact on the care of patients with locally advanced rectal cancer.Disclosure of Conflicts of Interest: D.M.K. disclosed no relevant relationships.References1. van der Sande ME, Beets GL, Hupkens BJ, et al. Response assessment after (chemo)radiotherapy for rectal cancer: Why are we missing complete responses with MRI and endoscopy? Eur J Surg Oncol 2019;45(6):1011–1017. Crossref, Medline, Google Scholar2. Maas M, Lambregts DMJ, Nelemans PJ, et al. Assessment of clinical complete response after chemoradiation for rectal cancer with digital rectal examination, endoscopy, and MRI: selection for organ-saving treatment. Ann Surg Oncol 2015;22(12):3873–3880. Crossref, Medline, Google Scholar3. Patel UB, Taylor F, Blomqvist L, et al. Magnetic resonance imaging-detected tumor response for locally advanced rectal cancer predicts survival outcomes: MERCURY experience. J Clin Oncol 2011;29(28):3753–3760. Crossref, Medline, Google Scholar4. Huh JW, Kim HC, Kim SH, et al. Tumor regression grade as a clinically useful outcome predictor in patients with rectal cancer after preoperative chemoradiotherapy. Surgery 2019;165(3):579–585. Crossref, Medline, Google Scholar5. Jang JK, Choi SH, Park SH, et al. MR tumor regression grade for pathological complete response in rectal cancer post neoadjuvant chemoradiotherapy: a systematic review and meta-analysis for accuracy. Eur Radiol 2020 Jan 17 [Epub ahead of print] https://doi.org/10.1007/s00330-019-06565-2. Crossref, Google Scholar6. Joye I, Deroose CM, Vandecaveye V, Haustermans K. The role of diffusion-weighted MRI and (18)F-FDG PET/CT in the prediction of pathologic complete response after radiochemotherapy for rectal cancer: a systematic review. Radiother Oncol 2014;113(2):158–165. Crossref, Medline, Google Scholar7. Rosenkrantz AB, Padhani AR, Chenevert TL, et al. Body diffusion kurtosis imaging: Basic principles, applications, and considerations for clinical practice. J Magn Reson Imaging 2015;42(5):1190–1202. Crossref, Medline, Google Scholar8. Hu F, Tang W, Sun Y, et al. The value of diffusion kurtosis imaging in assessing pathological complete response to neoadjuvant chemoradiation therapy in rectal cancer: a comparison with conventional diffusion-weighted imaging. Oncotarget 2017;8(43):75597–75606. Crossref, Medline, Google Scholar9. Zhang XY, Wang L, Zhu HT, et al. Predicting rectal cancer response to neoadjuvant chemoradiotherapy using deep learning of diffusion kurtosis MRI. Radiology 2020;296:56–64. Link, Google ScholarArticle HistoryReceived: Feb 11 2020Revision received: Feb 17 2020Revision received: Feb 25 2020Accepted: Feb 27 2020Published online: Apr 21 2020Published in print: July 2020 FiguresReferencesRelatedDetailsCited ByApplication of artificial intelligence in diagnosis and treatment of colorectal cancer: A novel ProspectZugangYin, ChenhuiYao, LiminZhang, ShaohuaQi2023 | Frontiers in Medicine, Vol. 10Deep Learning-Based Multimodal 3 T MRI for the Diagnosis of Knee OsteoarthritisYongHu, JieTang, ShenghaoZhao, YeLi, Ahmed FaeqHussein2022 | Computational and Mathematical Methods in Medicine, Vol. 2022Are We There Yet? The Value of Deep Learning in a Multicenter Setting for Response Prediction of Locally Advanced Rectal Cancer to Neoadjuvant ChemoradiotherapyBarbara D.Wichtmann, SteffenAlbert, WenzhaoZhao, AngelikaMaurer, ClausRödel, Ralf-DieterHofheinz, JürgenHesser, Frank G.Zöllner, Ulrike I.Attenberger2022 | Diagnostics, Vol. 12, No. 7Toward Generalizability in the Deployment of Artificial Intelligence in Radiology: Role of Computation Stress Testing to Overcome UnderspecificationThomas Eche, Lawrence H. Schwartz, Fatima-Zohra Mokrane, Laurent Dercle, 27 October 2021 | Radiology: Artificial Intelligence, Vol. 3, No. 6Accompanying This ArticlePredicting Rectal Cancer Response to Neoadjuvant Chemoradiotherapy Using Deep Learning of Diffusion Kurtosis MRIApr 21 2020RadiologyRecommended Articles Predicting Rectal Cancer Response to Neoadjuvant Chemoradiotherapy Using Deep Learning of Diffusion Kurtosis MRIRadiology2020Volume: 296Issue: 1pp. 56-64Watch-and-Wait Approach to Rectal Cancer: The Role of ImagingRadiology2023Volume: 307Issue: 1MRI for T Restaging of Locally Advanced Rectal Cancer Following Neoadjuvant Chemotherapy and Radiation TherapyRadiology2022Volume: 305Issue: 2pp. 373-374An MRI Deep Learning Model Predicts Outcome in Rectal CancerRadiology2023Volume: 307Issue: 5High-Precision Assessment of Chemoradiotherapy of Rectal Cancer with Near-Infrared Photoacoustic Microscopy and Deep LearningRadiology2021Volume: 299Issue: 2pp. 359-361See More RSNA Education Exhibits MRI Evaluation of Pathological Complete Response of Locally Advanced Rectal Cancer after Neoadjuvant Chemoradiation Therapy: Can We Wait and What Will We Watch?Digital Posters2019Case Based Review: MRI Assessment Of Rectal Carcinoma After Neoadjuvant Chemoradiation TherapyDigital Posters2021Extramural Venous Invasion in Rectal Cancer Before and After Neoadjuvant Therapy Beyond Classic Findings: Recognition of Detailed Characteristics and Tumor DepositDigital Posters2022 RSNA Case Collection Hydrogel infiltration into rectal wallRSNA Case Collection2020Rectal EndometriosisRSNA Case Collection2021Pseudoprogression with Immunotherapy Treatment RSNA Case Collection2021 Vol. 296, No. 1 Metrics Altmetric Score PDF download
Referência(s)