RPS 705 - Radiomics applications in MRI

Author Block: J-Y. Lin¹, C. Lu², Y-S. Hu³, J. K. Loo¹, K. L. Lee¹, C. Liao⁴, C-J. Lin¹; ¹Taipei/TW, ²Taipei City/TW, ³New Taipei/TW, ⁴Texas, TX/US
Purpose: Angioarchitectural analysis of brain arteriovenous malformations (BAVMs) is qualitative and subject to interpretation. This study quantified the morphology of and signal changes in the nidal and perinidal areas by using MR radiomics and compared the performance of MR radiomics and angioarchitectural analysis in detecting epileptic BAVMs.
Methods or Background: From 2010 to 2020, a total of 111 patients with supratentorial BAVMs were retrospectively included and grouped in accordance with the initial presentation of seizure. Patients' angiograms and MR imaging results were analysed to determine the corresponding angioarchitecture. The BAVM nidus was contoured on time-of-flight MR angiography images. The perinidal brain parenchyma was contoured on T2-weighted images, followed by radiomic analysis. Logistic regression analysis was performed to determine the independent risk factors for seizure. ROC curve analysis, decision curve analysis (DCA), and calibration curve were performed to compare the performance of angioarchitecture-based and radiomics-based models in diagnosing epileptic BAVMs.
Results or Findings: In multivariate analyses, low sphericity (OR: 2012.07, p=0.04) and angiogenesis (OR: 5.30, p=0.01) were independently associated with a high risk of seizure after adjustment for age, sex, temporal location, and nidal volume. The AUC for the angioarchitecture-based, MR radiomics-based, and combined models was 0.672, 0.817, and 0.794, respectively. DCA confirmed the clinical utility of the MR radiomics-based and combined models.
Conclusion: Low nidal sphericity and angiogenesis were associated with high seizure risk in patients with BAVMs. MR radiomics-derived tools may be used for noninvasive and objective measurement for evaluating the risk of seizure due to BAVM.
Limitations: Although the dataset acquired from a single institution and machine may improve the homogeneity of image data quality, an external validation dataset should be considered in future studies to improve the generalisability of prediction models.
Funding for this study: Funding was received from the Taipei Veterans General Hospital (grant number: V111C- 073) and Taiwan’s Ministry of Science and Technology (grant number: MOST- 109–2628-B-0).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the Third Institutional Review Board of Taipei Veterans General Hospital. The protocol was implemented after review and approval by the Human Research Protection Center of TPEVGH (IRB No.: 2020–06-005C).

Author Block: V. D'Agostino¹, M. P. Aparisi Gomez², R. Sassi¹, A. G. Morganti¹, M. Buwenge¹, A. Bazzocchi¹; ¹Bologna/IT, ²Auckland/NZ
Purpose: Pain management of bone metastases is performed with systemic and local therapies. External Beam radiotherapy is currently the gold standard for treatment of painful metastases, however MRI-guided high-intensity focused ultrasound (MRgHIFU) has shown great results in pain relief. To date, a reliable imaging method to predict the success of the treatment is yet to be defined. This work aims to investigate the potential role of a radiomics-based machine-learning (ML) algorithm applied to pre- and post-treatment MR T1w and T2w-images for the prediction of a clinical success (reduction of ≥6 point in the numerical rating pain scale) of MRgHIFU treatment.
Methods or Background: 188 patients (112 females, 76 males) with 200 bone metastases, who underwent MRgHIFU ablation were retrospectively selected and classified into two groups, on the basis of clinical success of the treatment. Two-dimensional segmentations were manually drawn by an MSK-expert radiologist on axial pre- and post-treatment T1w and T2w sequences. Radiomic feature extraction was performed using PyRadiomics. To reduce dimensionality, variance and intercorrelation analysis were used. Subsequently, a LogitBoost classifier was trained with stratified cross-validation, tested and validated within our population.
Results or Findings: Group A (Responders) reported 112 lesions; group B (Not-responders) reported 88 lesions. 3567 radiomics features were extracted, of which 2864 were discarded due to high intercorrelation (>0.8). The feature selection process identified ten features to build the ML classifier, which was able to correctly classify 94% of instances on the training set and 85% on the testing set. Weighted average precision and recall were 0.90 and 0.92 respectively, while the AUROC curve was 0.88. The performance was similar within the validation set.
Conclusion: An ML-classifier powered by MRI radiomics might be a feasible tool to predict bone metastases pain response to MRgHIFU.
Limitations: The study was limited by a lack of external validation, mild class imbalance, and its retrospective nature.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: No information provided by the submitter.

Author Block: M. Wennmann, L. Rotkopf, F. Bauer, H. Goldschmidt, T. F. Weber, H-P. Schlemmer, S. Delorme, K. H. Maier-Hein, P. Neher; Heidelberg/DE
Purpose: The aim of this study was to evaluate the influence of using only a subset of reproducible radiomics features, defined in a prior in-vivo multi-MRI-scanner test-retest-study, on the generalisability and external performance of radiomics models.
Methods or Background: This retrospective study used data acquired between 2015 and 2021. The task for the radiomics models was to predict bone marrow plasma cell infiltration from MRI in myeloma patients. Different machine learning (ML) models were trained on data from Centre 1, using either all radiomics features, or using only reproducible radiomics features defined by a prior in-vivo multi-MRI-scanner test-retest study. Models were tested on an internal and a multicentric external data set. Pearson correlation coefficient r and mean absolute error (MAE) between predicted and actual plasma cell infiltration were used to quantify the model performance. The difference between performance on the internal and external test set was calculated to measure generalisability.
Results or Findings: 302 MRIs from 300 patients from 8 centres were included. When using only reproducible features compared to all features, for all ML models the generalisability improved. However, for the best model, a random forest regressor, the model using all features still outperformed the model using only reproducible features on the external test set (r of 0.44 vs. 0.33 and MAE of 20.5 vs 21.9). When comparing the external performance across all combinations of ML models and feature selection methods, a random forest regressor using all features (r=0.44, MAE=20.5) showed the best external performance.
Conclusion: A radiomics feature selection based on in-vivo reproducibility experiments between different MRI scanners improves the generalisability of radiomics models, however, does not necessarily lead to an improvement of the external performance of the overall best radiomics model.
Limitations: The study is limited by its retrospective nature.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the local Institutional Research Board with code: S-537/2020.

Author Block: T. Hu¹, J. Gong¹, Y. Sun¹, M. Li¹, C. Cai¹, Y. Cui², X. Zhang³, T. Tong¹, Y. Feng¹; ¹Shanghai/CN, ²Taiyuan/CN, ³Beijing/CN
Purpose: The aim of this study was to investigate the ability of the MRI-based radiomics models for pretreatment prediction of good response (GR) to neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer (LARC).
Methods or Background: A total of 921 patients with LARC were retrospectively recruited from 3 hospitals, including a training dataset (TD) (n=508) and external validation datasets 1 (EVD1) (n=242) and 2 (EVD2) (n=171). Radiomics features were extracted from the T2WI and ADC images. Three classifications, including logistic regression (LR), random forest (RF), and support vector machine (SVM) were applied to construct radiomics models for predicting GR. The clinical-MRI model was constructed with significant clinical characteristics and MRI morphological features by using the logistic regression analysis. The prediction performance was evaluated using the area under the curve (AUC) and decision curve analysis (DCA).
Results or Findings: Two clinical-MRI features and ten radiomic features were selected for the GR prediction. Compared to models from other classifiers and the clinical-MRI model, the model obtained with SVM showed promising discrimination of GR to nCRT with AUCs of 0.798 (95% CI, 0.758-0.837), 0.790 (95% CI, 0.725-0.856) and 0.743 (95% CI, 0.666-0.821) in the training and two external validation datasets respectively. Decision curve analysis confirmed that the radiomics models were clinically useful.
Conclusion: The MRI-based radiomics model exhibited better performance for response prediction to nCRT in LARC patients than the clinical-MRI model, and also provided value for prognosis prediction.
Limitations: Selection bias may have been introduced by excluding patients with clinical complete response. The ROI did not include lymph nodes. Manual segmentation of ROIs is a time-consuming procedure and requires accurate identification of MRI lesions.
Funding for this study: Funding for this study was received from the National Natural Science Foundation of China (No.82001776, 81971687, 82271946),Shanghai Natural Science Foundation (No.20ZR1412700).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by the Institutional Review Boards of all participating centres, and the requirement for informed consent was waived due to the retrospective nature.

Author Block: W. Chen, G. Lin, J. Ji; Lishui/CN
Purpose: The aim of this study was to develop and validate a machine learning-based fusion model to preoperatively predict Ki-67 expression levels in patients with head and neck squamous cell carcinoma (HNSCC) using multiparametric magnetic resonance imaging (MRI) images.
Methods or Background: A total of 152 patients with pathologically proven HNSCC were retrospectively enrolled and divided into training (n=106) and validation (n=46) cohorts. Features were extracted from T2-weighted imaging fat suppression and contrast-enhanced T1-weighted images and screened using the least absolute shrinkage and selection operator (LASSO) regression. Seven machine learning classifiers, including k-nearest neighbors (KNN), support vector machine (SVM), logistic regression (LR), random forest (RF), linear discriminant analysis (LDA), naive Bayes (NB), and eXtreme Gradient Boosting (XGBoost) were trained. The best classifier was evaluated to calculate radiomics (Rad)-scores and combined clinical factors to build a fusion model, which was visualised as a nomogram. Performance was evaluated by the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis (DCA).
Results or Findings: The SVM classifier showed the best performance, with an AUC of 0.862 in the validation cohort. The fusion model incorporating SVM-based Rad-scores, clinical T stage and MR-reported LN status was constructed and achieved AUCs of 0.915 (0.864–0.967) and 0.896 (0.770–0.966) in the training and validation cohorts, with accuracies of 91.98% and 84.78%, respectively. The calibration curves demonstrated a good model fit, and DCA showed the clinical benefits of the fusion model.
Conclusion: The machine learning-based fusion model based on multiparametric MRI can predict the expression of Ki-67 in HNSCC patients, which might be helpful for prognosis evaluation and clinical decision-making.
Limitations: This is a retrospective single-centre study with unavoidable bias and a limited sample size.
Funding for this study: Funding for this study was received from the National Natural Science Foundation of China (No.82072026).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. This study was approved by the Institutional Review Board and Human Ethics Committee of the Fifth Affiliated Hospital of Wenzhou Medical University (No. 2023-523), with the requirement for patient informed consent being waived due to its retrospective nature. All patients’ information was anonymised prior to the analysis.

Author Block: I. Mariani, C. Maino, T. P. Giandola, C. Talei Franzesi, D. Ippolito; Monza/IT
Purpose: The aim of this study was to retrospectively collect radiomic data from preoperative rectal MR and determine the possible relationships between texture analysis and response to neoadjuvant treatment.
Methods or Background: 88 patients with biopsy-proven advanced rectal adenocarcinoma, staging MR and RAR after neoadjuvant treatmente were enrolled. Based on tumour regression grade, we considered TRG 1-2 patients as responders and TRG 3-5 patients as non-responders. Texture analysis was conducted by using LIFex software, where T2-weighted oblique axial MR sequences were uploaded; a region-of-interest (ROI) was manually drawn on a single slice. Features with a Spearman correlation index >0.5 have been discarded and a LASSO feature selection has been applied. Selected features were trained using bootstrapping.
Results or Findings: According to TRG classes 49 patients (55.8%) were considered responders while 39 (44.2%) as non-responders. Two features were associated with responders’ classes: GLCM_Homogeneity and Discretised Histo Entropy log 2. Regarding GLCM_Homogeneity, the area under the receiver operating characteristic curve (AUC), sensitivity (sens), specificity (spec), positive predictive value (PPV), and negative predictive value (NPV) were: 0.779 (95% CIs=0.771-0.816), 86% (80-90), 67% (60-71%), 81% (76-84), and 88% (84-90), respectively. Regarding Discretised Histo Entropy log 2, diagnostic values were as follows: AUC=0.775 (0.700-0.801), sens=80% (74-83), spec=63% (58-69%), PPV=77% (70-81), and NPV=82% (80-85). By combing both radiomics features the radiomics signature diagnostic accuracy increased (AUC=0.844, p<0.05). Finally, the AUC of 1000 bootstraps was 0.810.
Conclusion: Texture analysis can be considered an advanced complementary diagnostic tool to determine a possible correlation between pre-surgical MR data and response to neoadjuvant therapy.
Limitations: Considering its low robustness, further studies with a larger cohort of patients should aim to validate these preliminary data.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: No information provided by the submitter.

Author Block: J. Lozano¹, A. Jimenez-Pastor¹, G. Weiss², D. Veiga Canuto¹, B. Martínez De Las Heras¹, A. Cañete Nieto¹, B. Hero³, R. Ladenstein⁴, L. Marti-Bonmati¹; ¹Valencia/ES, ²Boston, MA/US, ³Cologne/DE, ⁴Vienna/AT
Purpose: Neuroblastoma (NB) is the most frequent and highly aggressive solid cancer in childhood, in which imaging plays a pivotal role at every step of the patient's journey. This study sought to develop a machine learning model using clinical, molecular, and magnetic resonance (MR) radiomics features to predict patient’s overall survival (OS) and aid in their stratification.
Methods or Background: A database of 513 patients was used for model training, validation, and testing. Furthermore, 22 additional patients from hospitals not originally in the database were utilised as an external test. Manual tumour segmentations of the NB were conducted on the corresponding T2-weighted MR images to segment the primary tumour by an experienced radiologist. In total, 107 radiomics features were extracted and subsequently harmonised across manufacturers and magnetic field strengths using the nested ComBat methodology. Finally, radiomic features were combined with the clinical and molecular data to serve as input for the models. A nested cross-validation approach was used as training methodology to select the best preprocessing and model configuration.
Results or Findings: A C-index of 0.788±0.049 was achieved in the test, being a random survival forest the model showing the best performance. For the additional 22 patients, a C-index of 0.934 was obtained. The model exhibited superior predictive performance and patient stratification compared to the standard risk group INRG. Interpretability analysis revealed the significance of clinical variables, with radiomics features related to lesion heterogeneity and size playing an important role in prediction.
Conclusion: The OS predictive model demonstrated high performance and alignment with established clinical variables, highlighting the importance of radiomics features. It presents new evidence for enhancing patient care and clinical decision-making.
Limitations: Greater sample sizes are required in the external test to confirm the results.
Funding for this study: Funding was received from PRIMAGE (PRedictive In-silico Multiscale Analytics to support cancer personalized diagnosis and prognosis, empowered by imaging biomarkers), a Horizon 2020|RIA project (Topic SC1-DTH-07-2018), grant agreement no: 826494.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by an Institutional Review Board and written informed consent was obtained from all participant centres.

Author Block: S. Long, J. Chen, L. Zhong, W. Liu, Y. Pei, W. Li; Changsha/CN
Purpose: The massive presence of peri-tumour tertiary lymphoid structure (mpTLS) questions whether HCC patients can benefit from immunotherapy. However, it has been identified only by pathological examination. This study aims to develop a noninvasive tool using preoperative multiple parameter MR imaging (mpMRI) radiomic for predicting mpTLS.
Methods or Background: 584 consecutive HCC patients (mpTLS+:154; mpTLS-:430) were retrospectively recruited from four independent institutes and were divided into training (n=307) in one institute, and a validation cohort (n=277) in the other three institute. 76 HCC participants (mpTLS+:21; mpTLS-:55) were also enrolled as prospective cohort, including the relapse-free survival (RFS) and overall survival (OS). All subjects underwent preoperative mpMRI. Three different models (Model 1: peri-tumour model; Model 2: intra-tumour model; Model 3: combined models 1 and 2) were constructed to stratify mpTLS+. The optimal model was decided by the maximum area under of curve (AUC) in the training set and validated in both the validation and the prospective cohort, which was further used to predict the RFS and OS in the prospective data.
Results or Findings: For retrospective data, Model 3 (AUC:0.92) was the optimal model for diagnosing mpTLS+ than model 2 (AUC:0.87) and model 1 (AUC:0.85) in training cohorts (all P<0.001), which was validated in validation cohort (AUC: 0.91 vs 0.85 vs 0.84; all P<0.001), and prospective corhort (AUC:0.91 vs 0.84 vs 0.77; all P<0.001). For prospective data, Model 3 could predict the RFS (P < 0.001) and OS (P<0.001) based on mpTLS+.
Conclusion: Model 3 (the combined model) is a reliable and noninvasive tool for predicting mpTLS and further can forecast OS and RFS, which is helpful in deciding immunotherapy for HCC patients.
Limitations: No limitations were identified.
Funding for this study: Funding was received from the National Natural Science Foundation of China [82071895 and 82271984] to W.Z.L.; Hunan Provincial Science and Technology Department [2023JJ30903 and 2022JJ30950] to W.Z.L. and Y.G.P. National Geriatric Disease Clinical Medical Research Center Foundation [2022LNJJ08] to Y.G.P. Youth Project of Natural Science Foundation of Hunan Province to W.G.L.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved with the Ethics Committee number: 2018111101.