Can Artificial Intelligence Fix the Reproducibility Problem of Radiomics?
2019; Radiological Society of North America; Volume: 292; Issue: 2 Linguagem: Inglês
10.1148/radiol.2019191154
ISSN1527-1315
Autores Tópico(s)Advanced X-ray and CT Imaging
ResumoHomeRadiologyVol. 292, No. 2 PreviousNext Reviews and CommentaryFree AccessEditorialCan Artificial Intelligence Fix the Reproducibility Problem of Radiomics?Chang Min Park Chang Min Park Author AffiliationsFrom the Department of Radiology, Seoul National University Hospital, Seoul National University College of Medicine, 101 Daehak-ro, Jongno-gu, Seoul 03080, Korea.Address correspondence to the author (e-mail: [email protected]).Chang Min Park Published Online:Jun 18 2019https://doi.org/10.1148/radiol.2019191154MoreSectionsPDF ToolsImage ViewerAdd to favoritesCiteTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinked In See also the article by Choe et al in this issue.IntroductionRadiomics is at the forefront of oncologic radiology research today (1). Radiomics converts medical images such as CT scans into high-throughput quantitative data that can be used to improve diagnostic, prognostic, and predictive accuracy. Indeed, radiomics can provide objective and comprehensive information regarding whole or subregions of cancers in a noninvasive and repeatable way (1). Radiomic signatures have been demonstrated to reflect intratumoral heterogeneity and to be associated with gene-expression profiles, both of which can serve as important prognostic factors (2). A myriad of research articles have showcased the potential of radiomics in providing useful imaging biomarkers in patients with cancer (1–3). However, radiomics has not yet been used in clinical practice due to several major obstacles, including: (a) lack of an easy-to-handle, high-performing set of analytic tools, and (b) variability of radiomic features that are prone to substantial influence from image acquisition parameters, as well as target selection, target segmentation, feature extraction/selection, and radiomics modeling.For CT radiomics, image reconstruction algorithms (ie, reconstruction kernels) and section thickness have been major sources of radiomic feature variability (4,5). Indeed, variability stemming from the use of different reconstruction kernels impedes the comparability of radiomic features between studies performed by different groups as well as in longitudinal studies that include a variety of imaging parameters. Theoretically, we may solve this issue by obtaining raw CT data (sinograms of CT acquisition) or by reconstructing and storing many combinations of CT reconstruction parameters. However, such solutions are not feasible in practice. Thus, another method to resolve this variability is warranted.In this issue of Radiology, Choe and colleagues (6) investigated the effect of different reconstruction kernels on CT radiomic features of lung nodules and attempted to improve the reproducibility of radiomic features by applying a convolutional neural network (CNN)–based kernel-conversion technique. To develop their conversion algorithm, they used residual learning in an end-to-end way. CT images reconstructed using a soft reconstruction kernel (termed B30f on their CT system) were converted by CNN to those of a sharp reconstruction kernel (B50f on their CT) and vice versa. The authors segmented 104 pulmonary nodules or masses and extracted 702 radiomic features from each lesion. Reproducibility of radiomic features was evaluated using the concordance correlation coefficient (CCC), which measures the agreement between two variables to evaluate reproducibility; CCC values of ± 1 denote perfect concordance and discordance, and a value of 0 denotes absence of concordance.Choe and colleagues found that the CCC between human readers for CT images with the same kernel setting was 0.92; the majority of radiomic features (84.3% of all 702 radiomic features) were reproducible (CCC ≥ 0.85). However, when it came to different reconstruction kernels, the story was different. The CCC between readers for images reconstructed with different CT kernels was markedly lower, at 0.38; only 15% (107 of 702 features) of the radiomic features were reproducible. Notably, texture and wavelet features were particularly influenced by the reconstruction kernels, both in non–contrast material–enhanced and contrast-enhanced CT scans. These results are in agreement with those of previous studies (4,5) and frankly were not surprising in the least.However, Choe et al reported impressive results when their CNN-based kernel-conversion algorithm was applied. The reproducibility of radiomic features using different reconstruction kernels showed dramatic improvement with their algorithm. In the pooled analysis of interreader variability using different kernels (B30f and B50f), CCCs improved from 0.38 to 0.84 after applying kernel conversion (CCC after images were converted from B50f to B30f, 0.84; CCC after images were converted from B30f to B50f, 0.83). The improvement in reproducibility was particularly noticeable in texture features and wavelet features, for which the reproducibility between the two kernels without kernel conversion was very low (CCCs of 0.61 and 0.35 for texture and wavelet features, respectively). Although we previously believed that different reconstruction kernels could not be used interchangeably in radiomics research, Choe and colleagues showed that it may be possible to compare radiomics features from CT images with different reconstruction kernels using this novel CNN-based kernel-conversion algorithm.Several issues remain to be solved prior to its application in real clinical practice. First, the actual effect of kernel conversion on the performance of radiomics was not analyzed. Thus, further studies must address whether the diagnostic, predictive, or prognostic values obtained using converted CT images would improve, preserve, or deteriorate its performance. Second, the study by Choe et al dealt with a very narrow and specific subject: kernel conversion between two reconstruction kernels (B30f and B50f) from a single CT scanner. It is possible, however, that variation between these soft and sharp kernels reflects the known variability between different CT manufacturers. Choe et al provided a clear example of the application of this brand-new and emerging technology (ie, CNN-based image style transfer). The impact of the authors' study is a major step forward in radiomics. Considering the rapid pace of deep learning technologies today, I am certain that a variety of deep learning–based solutions will eventually solve nearly every step of the radiomics process, including CT acquisition, target identification, segmentation, and feature extraction in the exciting years to come.In conclusion, a convolutional neural network–based kernel-conversion algorithm dramatically improved the similarity of CT radiomic features obtained using different reconstruction kernels. Deep learning–based approaches are expected to substantially contribute to the applicability of radiomics. The study by Choe et al will serve as one of the first steps toward bigger strides in radiomics.Disclosures of Conflicts of Interest: C.M.P. disclosed no relevant relationships.References1. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than pictures, they are data. Radiology 2016;278(2):563–577. Link, Google Scholar2. Aerts HJ, Velazquez ER, Leijenaar RT, et al. Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach. Nat Commun 2014;5(1):4006. Crossref, Medline, Google Scholar3. Lambin P, Leijenaar RTH, Deist TM, et al. Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 2017;14(12):749–762. Crossref, Medline, Google Scholar4. Kim H, Park CM, Lee M, et al. Impact of reconstruction algorithms on CT radiomic features of pulmonary tumors: analysis of intra- and inter-reader variability and inter-reconstruction algorithm variability. PLoS One 2016;11(10):e0164924. Crossref, Medline, Google Scholar5. Berenguer R, Pastor-Juan MDR, Canales-Vázquez J, et al. Radiomics of CT features may be nonreproducible and redundant: influence of CT acquisition parameters. Radiology 2018;288(2):407–415. Link, Google Scholar6. Choe J, Lee SM, Do KH, et al. Deep learning-based image conversion of CT reconstruction kernels improves radiomics reproducibility for pulmonary nodules or masses. Radiology 2019;292;365–373. Link, Google ScholarArticle HistoryReceived: May 21 2019Revision requested: May 28 2019Revision received: May 28 2019Accepted: May 29 2019Published online: June 18 2019Published in print: Aug 2019 FiguresReferencesRelatedDetailsCited ByGenerative adversarial network with radiomic feature reproducibility analysis for computed tomography denoisingJinaLee, JaeikJeon, YoungtaekHong, DawunJeong, YeonggulJang, ByunghwanJeon, Hye JinBaek, EunCho, HackjoonShim, Hyuk-JaeChang2023 | Computers in Biology and Medicine, Vol. 159Annotation-Efficient Deep Learning Model for Pancreatic Cancer Diagnosis and Classification Using CT Images: A Retrospective Diagnostic StudyThanapornViriyasaranon, Jung WonChun, Young HwanKoh, Jae HeeCho, Min KyuJung, Seong-HunKim, Hyo JungKim, Woo JinLee, Jang-HwanChoi, Sang MyungWoo2023 | Cancers, Vol. 15, No. 13Assessing the predictability of the H3K27M status in diffuse glioma patients using frequency importance analysis on chemical exchange saturation transfer MRIYibingChen, BenqiZhao, ChanghaoZhu, ChongxueBie, XiaoweiHe, ZhuozhaoZheng, XiaoleiSong2023 | Magnetic Resonance Imaging, Vol. 103Deep Radiomics–based Approach to the Diagnosis of Osteoporosis Using Hip RadiographsSangwook Kim, Bo Ram Kim, Hee-Dong Chae, Jimin Lee, Sung-Joon Ye, Dong Hyun Kim, Sung Hwan Hong, Ja-Young Choi, Hye Jin Yoo, 25 May 2022 | Radiology: Artificial Intelligence, Vol. 4, No. 4A meta-analysis of the diagnostic test accuracy of CT-based radiomics for the prediction of COVID-19 severityYung-ShuoKao, Kun-TeLin2022 | La radiologia medica, Vol. 127, No. 718F-FDG PET Radiomics as Predictor of Treatment Response in Oesophageal Cancer: A Systematic Review and Meta-AnalysisLetiziaDeantonio, Maria LuisaGaro, GaetanoPaone, Maria CarlaValli, StefanoCappio, DavideLa Regina, MarcoCefali, Maria CelestePalmarocchi, AlbertoVannelli, SaraDe Dosso2022 | Frontiers in Oncology, Vol. 12Radiomics for Predicting Response of Neoadjuvant Chemotherapy in Nasopharyngeal Carcinoma: A Systematic Review and Meta-AnalysisChaoYang, ZekunJiang, TingtingCheng, RongrongZhou, GuangcanWang, DiJing, LinlinBo, PuHuang, JianboWang, DaizhouZhang, JianweiJiang, XingWang, HuaLu, ZijianZhang, DengwangLi2022 | Frontiers in Oncology, Vol. 12Radiomics-Based Deep Learning Prediction of Overall Survival in Non-Small-Cell Lung Cancer Using Contrast-Enhanced Computed TomographyKuei-YuanHou, Jyun-RuChen, Yung-ChenWang, Ming-HuangChiu, Sen-PingLin, Yuan-HengMo, Shih-ChiehPeng, Chia-FengLu2022 | Cancers, Vol. 14, No. 15Enhanced Hierarchical Feature Synthesis Network for the Improvement of Computed Tomography Radiomic Features ReproducibilityDawunJeong, YoungtaekHong, JinaLee, Seul BiLee, Yeon JinCho2022 | SSRN Electronic JournalUsing deep learning to predict microvascular invasion in hepatocellular carcinoma based on dynamic contrast-enhanced MRI combined with clinical parametersDanjunSong, YueyueWang, WentaoWang, YiningWang, JiabinCai, KaiZhu, MinzhiLv, QiangGao, JianZhou, JiaFan, ShengxiangRao, ManningWang, XiaoyingWang2021 | Journal of Cancer Research and Clinical Oncology, Vol. 147, No. 12Future artificial intelligence tools and perspectives in medicineAhmadChaddad, YousefKatib, LamaHassan2021 | Current Opinion in Urology, Vol. 31, No. 4Pulmonary Hypertension in Association with Lung Disease: Quantitative CT and Artificial Intelligence to the Rescue? State-of-the-Art ReviewKritDwivedi, MichaelSharkey, RobinCondliffe, Johanna M.Uthoff, SamerAlabed, PeterMetherall, HaipingLu, Jim M.Wild, Eric A.Hoffman, Andrew J.Swift, David G.Kiely2021 | Diagnostics, Vol. 11, No. 4Decoding and Systematization of Medical Imaging Features of Multiple Human MalignanciesLu Wang, Zhaoyu Liu, Jiayi Xie, Yuheng Chen, Xiaoqi Zhao, Zifan You, Mingshu Yang, Wei Qian, Jie Tian, Kristen Yeom, Jiangdian Song, 11 September 2020 | Radiology: Imaging Cancer, Vol. 2, No. 5Radiomics feature reproducibility under inter-rater variability in segmentations of CT imagesChristophHaarburger, GustavMüller-Franzes, LeonWeninger, ChristianeKuhl, DanielTruhn, DoritMerhof2020 | Scientific Reports, Vol. 10, No. 1Current and Potential Applications of Artificial Intelligence in Gastrointestinal Stromal Tumor ImagingCai-WeiYang, Xi-JiaoLiu, Si-YunLiu, ShangWan, ZhengYe, BinSong, Jian-GuoZhou2020 | Contrast Media & Molecular Imaging, Vol. 2020Accompanying This ArticleDeep Learning–based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or MassesJun 18 2019RadiologyRecommended Articles Deep Learning–based Image Conversion of CT Reconstruction Kernels Improves Radiomics Reproducibility for Pulmonary Nodules or MassesRadiology2019Volume: 292Issue: 2pp. 365-373Deep Learning: A Primer for RadiologistsRadioGraphics2017Volume: 37Issue: 7pp. 2113-2131Current Applications and Future Impact of Machine Learning in RadiologyRadiology2018Volume: 288Issue: 2pp. 318-328Photon-counting Detector CT with Deep Learning Noise Reduction to Detect Multiple MyelomaRadiology2022Volume: 306Issue: 1pp. 229-236Deep Learning–based Reconstruction for Lower-Dose Pediatric CT: Technical Principles, Image Characteristics, and Clinical ImplementationsRadioGraphics2021Volume: 41Issue: 7pp. 1936-1953See More RSNA Education Exhibits Ultra-high-resolution CT With 1024-matrix Using A Novel Deep Learning Reconstruction Technique: How To Use It For Abdominal ImagingDigital Posters2021Artificial Intelligence in Diagnostic Imaging: Current Applications and Future PerspectiveDigital Posters2019A Quick Start to Deep Learning for RadiologistsDigital Posters2019 RSNA Case Collection Stroke following CT guided chest biopsy RSNA Case Collection2020Diffuse idiopathic pulmonary neuroendocrine cell hyperplasiaRSNA Case Collection2020Nodular Pulmonary AmyloidosisRSNA Case Collection2021 Vol. 292, No. 2 Metrics Altmetric Score PDF download
Referência(s)