The Image Biomarker Standardization Initiative: Standardized Convolutional Filters for Reproducible Radiomics and Enhanced Clinical Insights

Filters are commonly used to enhance specific structures and patterns in images, such as vessels or peritumoral regions, to enable clinical insights beyond the visible image using radiomics. However, their lack of standardization restricts reproducibility and clinical translation of radiomics decision support tools. In this special report, teams of researchers who developed radiomics software participated in a three-phase study (September 2020 to December 2022) to establish a standardized set of filters. The first two phases focused on finding reference filtered images and reference feature values for commonly used convolutional filters: mean, Laplacian of Gaussian, Laws and Gabor kernels, separable and nonseparable wavelets (including decomposed forms), and Riesz transformations. In the first phase, 15 teams used digital phantoms to establish 33 reference filtered images of 36 filter configurations. In phase 2, 11 teams used a chest CT image to derive reference values for 323 of 396 features computed from filtered images using 22 filter and image processing configurations. Reference filtered images and feature values for Riesz transformations were not established. Reproducibility of standardized convolutional filters was validated on a public data set of multimodal imaging (CT, fluorodeoxyglucose PET, and T1-weighted MRI) in 51 patients with soft-tissue sarcoma. At validation, reproducibility of 486 features computed from filtered images using nine configurations × three imaging modalities was assessed using the lower bounds of 95% CIs of intraclass correlation coefficients. Out of 486 features, 458 were found to be reproducible across nine teams with lower bounds of 95% CIs of intraclass correlation coefficients greater than 0.75. In conclusion, eight filter types were standardized with reference filtered images and reference feature values for verifying and calibrating radiomics software packages. A web-based tool is available for compliance checking.

Tensor radiomics: paradigm for systematic incorporation of multi-flavoured radiomics features

Background

Radiomics features hold significant value as quantitative imaging biomarkers for diagnosis, prognosis, and treatment response assessment. To generate radiomics features and ultimately develop signatures, various factors can be manipulated, including image discretization parameters (e.g., bin number or size), convolutional filters, segmentation perturbation, or multi-modality fusion levels. Typically, only one set of parameters is employed, resulting in a single value or "flavour" for each radiomics feature. In contrast, we propose "tensor radiomics" (TR) where tensors of features calculated using multiple parameter combinations (i.e., flavours) are utilized to optimize the creation of radiomics signatures.

Methods

We provide illustrative instances of TR implementation in positron emission tomography-computed tomography (PET-CT), magnetic resonance imaging (MRI), and CT by leveraging machine learning (ML) and deep learning (DL) methodologies, as well as reproducibility analyses: (I) to predict overall survival (OS) in lung cancer (CT) and head and neck cancer (PET-CT), TR was employed by varying bin sizes. This approach involved use of a hybrid deep neural network called 'TR-Net' and two ML-based techniques for combining different flavours. (II) TR was constructed by incorporating different segmentation perturbations and various bin sizes to classify the response of late-stage lung cancer to first-line immunotherapy using CT images. (III) In MRI of glioblastoma (GBM), TR was implemented to generate multi-flavour radiomics features, enabling enhanced analysis and interpretation. (IV) TR was employed via multiple PET-CT fusions in head and neck cancer. Flavours based on different fusions were created using Laplacian pyramids and wavelet transforms.

Results

Our findings demonstrated that TR outperformed conventional radiomics features in lung cancer CT and head and neck cancer PET-CT images, significantly enhancing OS prediction accuracy. TR also improved classification of lung cancer response to therapy and exhibited notable advantages in reproducibility compared to single-flavour features in MR imaging of GBM. Moreover, in head and neck cancer, TR through multiple PET-CT fusions exhibited improved performance in predicting OS.

Conclusions

We conclude that the proposed TR paradigm has significant potential to improve performance in different medical imaging tasks. By incorporating multiple flavours of radiomics features, TR overcomes limitations associated with individual features and shows promise in enhancing prognostic capabilities in clinical settings.

Fusion-based tensor radiomics using reproducible features: Application to survival prediction in head and neck cancer

BACKGROUND

Numerous features are commonly generated in radiomics applications as applied to medical imaging, and identification of robust radiomics features (RFs) can be an important step to derivation of reliable, reproducible solutions. In this work, we utilize a tensor radiomics (TR) framework, where numerous fusions are explored, to generate different flavours of RFs, and we aimed to identify RFs that are robust to fusion techniques in head and neck cancer. Overall, we aimed to predict progression-free survival (PFS) using Hybrid Machine Learning Systems (HMLS) and reproducible RFs.

METHODS

The study was performed on 408 patients with head and neck cancer from The Cancer Imaging Archive. After image preprocessing, 15 fusion techniques were employed to combine Positron Emission Tomography (PET) and Computed Tomography (CT) images. Subsequently, 215 RFs were extracted through a standardized radiomics software, with 17 'flavours' generated using PET-only, CT-only, and 15 fused PET&CT images. The variability of RFs across flavours was studied using the Intraclass Correlation Coefficient (ICC). Furthermore, the features were categorized into seven reliability groups, 106 reproducible RFs with ICC>0.75 were selected, highly correlated flavours were removed, Principal Component Analysis was used to convert 17 flavours to 1 attribute, the polynomial function was utilized to increase RFs, and Analysis of variance (ANOVA) was used to select the relevant attributes. Finally, 3 classifiers including Random Forest (RFC), Logistic regression (LR), and Multi-layer perceptron were applied to the preselected relevant attributes to predict binary PFS. In 5-fold cross-validation, 80% of 4 divisions were utilized to train the model, and the remaining 20% was utilized to evaluate the model. Further, the remaining fold was used for external nested testing.

RESULTS

Reliability analysis indicated that most morphological features belong to the high-reliability category. By contrast, local intensity and statistical features extracted from images belong to the low-reliability category. In the tensor framework, the highest 5-fold cross-validation accuracy of 76.7%±3.3% with an external nested testing of 70.6%±6.7% resulted from the reproducible TR+polynomial function+ANOVA+LR algorithm while the accuracy of 70.0%±4.2% with the external nested testing of 67.7%±4.9% was achieved through the PCA fusion+RFC (non-tensor paradigm).

CONCLUSIONS

This study demonstrated that using reproducible RFs as utilized within a tensor fusion radiomics framework, linked with ANOVA and LR, added value to prediction of progression-free survival outcome in head and neck cancer patients.

Prediction of Parkinson's disease pathogenic variants using hybrid Machine learning systems and radiomic features

PURPOSE

In Parkinson's disease (PD), 5-10% of cases are of genetic origin with mutations identified in several genes such as leucine-rich repeat kinase 2 (LRRK2) and glucocerebrosidase (GBA). We aim to predict these two gene mutations using hybrid machine learning systems (HMLS), via imaging and non-imaging data, with the long-term goal to predict conversion to active disease.

METHODS

We studied 264 and 129 patients with known LRRK2 and GBA mutations status from PPMI database. Each dataset includes 513 features such as clinical features (CFs), conventional imaging features (CIFs) and radiomic features (RFs) extracted from DAT-SPECT images. Features, normalized by Z-score, were univariately analyzed for statistical significance by the t-test and chi-square test, adjusted by Benjamini-Hochberg correction. Multiple HMLSs, including 11 features extraction (FEA) or 10 features selection algorithms (FSA) linked with 21 classifiers were utilized. We also employed Ensemble Voting (EV) to classify the genes.

RESULTS

For prediction of LRRK2 mutation status, a number of HMLSs resulted in accuracies of 0.98 ± 0.02 and 1.00 in 5-fold cross-validation (80% out of total data points) and external testing (remaining 20%), respectively. For predicting GBA mutation status, multiple HMLSs resulted in high accuracies of 0.90 ± 0.08 and 0.96 in 5-fold cross-validation and external testing, respectively. We additionally showed that SPECT-based RFs added value to the specific prediction of of GBA mutation status.

CONCLUSION

We demonstrated that combining medical information with SPECT-based imaging features, and optimal utilization of HMLS can produce excellent prediction of the mutations status in PD patients.

Deep versus Handcrafted Tensor Radiomics Features: Prediction of Survival in Head and Neck Cancer Using Machine Learning and Fusion Techniques

BACKGROUND

Although handcrafted radiomics features (RF) are commonly extracted via radiomics software, employing deep features (DF) extracted from deep learning (DL) algorithms merits significant investigation. Moreover, a "tensor'' radiomics paradigm where various flavours of a given feature are generated and explored can provide added value. We aimed to employ conventional and tensor DFs, and compare their outcome prediction performance to conventional and tensor RFs.

METHODS

408 patients with head and neck cancer were selected from TCIA. PET images were first registered to CT, enhanced, normalized, and cropped. We employed 15 image-level fusion techniques (e.g., dual tree complex wavelet transform (DTCWT)) to combine PET and CT images. Subsequently, 215 RFs were extracted from each tumor in 17 images (or flavours) including CT only, PET only, and 15 fused PET-CT images through the standardized-SERA radiomics software. Furthermore, a 3 dimensional autoencoder was used to extract DFs. To predict the binary progression-free-survival-outcome, first, an end-to-end CNN algorithm was employed. Subsequently, we applied conventional and tensor DFs vs. RFs as extracted from each image to three sole classifiers, namely multilayer perceptron (MLP), random-forest, and logistic regression (LR), linked with dimension reduction algorithms.

RESULTS

DTCWT fusion linked with CNN resulted in accuracies of 75.6 ± 7.0% and 63.4 ± 6.7% in five-fold cross-validation and external-nested-testing, respectively. For the tensor RF-framework, polynomial transform algorithms + analysis of variance feature selector (ANOVA) + LR enabled 76.67 ± 3.3% and 70.6 ± 6.7% in the mentioned tests. For the tensor DF framework, PCA + ANOVA + MLP arrived at 87.0 ± 3.5% and 85.3 ± 5.2% in both tests.

CONCLUSIONS

This study showed that tensor DF combined with proper machine learning approaches enhanced survival prediction performance compared to conventional DF, tensor and conventional RF, and end-to-end CNN frameworks.

Prediction of Cognitive Decline in Parkinson's Disease Using Clinical and DAT SPECT Imaging Features, and Hybrid Machine Learning Systems

BACKGROUND

We aimed to predict Montreal Cognitive Assessment (MoCA) scores in Parkinson's disease patients at year 4 using handcrafted radiomics (RF), deep (DF), and clinical (CF) features at year 0 (baseline) applied to hybrid machine learning systems (HMLSs).

METHODS

297 patients were selected from the Parkinson's Progressive Marker Initiative (PPMI) database. The standardized SERA radiomics software and a 3D encoder were employed to extract RFs and DFs from single-photon emission computed tomography (DAT-SPECT) images, respectively. The patients with MoCA scores over 26 were indicated as normal; otherwise, scores under 26 were indicated as abnormal. Moreover, we applied different combinations of feature sets to HMLSs, including the Analysis of Variance (ANOVA) feature selection, which was linked with eight classifiers, including Multi-Layer Perceptron (MLP), K-Neighbors Classifier (KNN), Extra Trees Classifier (ETC), and others. We employed 80% of the patients to select the best model in a 5-fold cross-validation process, and the remaining 20% were employed for hold-out testing.

RESULTS

For the sole usage of RFs and DFs, ANOVA and MLP resulted in averaged accuracies of 59 ± 3% and 65 ± 4% for 5-fold cross-validation, respectively, with hold-out testing accuracies of 59 ± 1% and 56 ± 2%, respectively. For sole CFs, a higher performance of 77 ± 8% for 5-fold cross-validation and a hold-out testing performance of 82 + 2% were obtained from ANOVA and ETC. RF+DF obtained a performance of 64 ± 7%, with a hold-out testing performance of 59 ± 2% through ANOVA and XGBC. Usage of CF+RF, CF+DF, and RF+DF+CF enabled the highest averaged accuracies of 78 ± 7%, 78 ± 9%, and 76 ± 8% for 5-fold cross-validation, and hold-out testing accuracies of 81 ± 2%, 82 ± 2%, and 83 ± 4%, respectively.

CONCLUSIONS

We demonstrated that CFs vitally contribute to predictive performance, and combining them with appropriate imaging features and HMLSs can result in the best prediction performance.

Application of novel hybrid machine learning systems and radiomics features for non-motor outcome prediction in Parkinson's disease

Objectives.Parkinson's disease (PD) is a complex neurodegenerative disorder, affecting 2%-3% of the elderly population. Montreal Cognitive Assessment (MoCA), a rapid nonmotor screening test, assesses different cognitive dysfunctionality aspects. Early MoCA prediction may facilitate better temporal therapy and disease control. Radiomics features (RF), in addition to clinical features (CF), are indicated to increase clinical diagnoses, etc, bridging between medical imaging procedures and personalized medicine. We investigate the effect of RFs, CFs, and conventional imaging features (CIF) to enhance prediction performance using hybrid machine learning systems (HMLS).Methods.We selected 210 patients with 981 features (CFs, CIFs, and RFs) from the Parkinson's Progression-Markers-Initiative database. We generated 4 datasets, namely using (i), (ii) year-0 (D1) or year-1 (D2) features, (iii) longitudinal data (D3, putting datasets in years 0 and 1 longitudinally next to each other), and (iv) timeless data (D4, effectively doubling dataset size by listing both datasets from years 0 and 1 separately). First, we directly applied 23 predictor algorithms (PA) to the datasets to predict year-4 MoCA, which PD patients this year have a higher dementia risk. Subsequently, HMLSs, including 14 attribute extraction and 10 feature selection algorithms followed by PAs were employed to enhance prediction performances. 80% of all datapoints were utilized to select the best model based on minimum mean absolute error (MAE) resulting from 5-fold cross-validation. Subsequently, the remaining 20% was used for hold-out testing of the selected models.Results.When applying PAs without ASAs/FEAs to datasets (MoCA outcome range: [11,30]), Adaboost achieved an MAE of 1.74 ± 0.29 on D4 with a hold-out testing performance of 1.71. When employing HMLSs, D4 + Minimum_Redundancy_Maximum_Relevance (MRMR)+K_Nearest_Neighbor Regressor achieved the highest performance of 1.05 ± 0.25 with a hold-out testing performance of 0.57.Conclusion.Our study shows the importance of using larger datasets (timeless), and utilizing optimized HMLSs, for significantly improved prediction of MoCA in PD patients.

Longitudinal clustering analysis and prediction of Parkinson's disease progression using radiomics and hybrid machine learning

BACKGROUND

We employed machine learning approaches to (I) determine distinct progression trajectories in Parkinson's disease (PD) (unsupervised clustering task), and (II) predict progression trajectories (supervised prediction task), from early (years 0 and 1) data, making use of clinical and imaging features.

METHODS

We studied PD-subjects derived from longitudinal datasets (years 0, 1, 2 & 4; Parkinson's Progressive Marker Initiative). We extracted and analyzed 981 features, including motor, non-motor, and radiomics features extracted for each region-of-interest (ROIs: left/right caudate and putamen) using our standardized standardized environment for radiomics analysis (SERA) radiomics software. Segmentation of ROIs on dopamine transposer - single photon emission computed tomography (DAT SPECT) images were performed via magnetic resonance images (MRI). After performing cross-sectional clustering on 885 subjects (original dataset) to identify disease subtypes, we identified optimal longitudinal trajectories using hybrid machine learning systems (HMLS), including principal component analysis (PCA) + K-Means algorithms (KMA) followed by Bayesian information criterion (BIC), Calinski-Harabatz criterion (CHC), and elbow criterion (EC). Subsequently, prediction of the identified trajectories from early year data was performed using multiple HMLSs including 16 Dimension Reduction Algorithms (DRA) and 10 classification algorithms.

RESULTS

We identified 3 distinct progression trajectories. Hotelling's t squared test (HTST) showed that the identified trajectories were distinct. The trajectories included those with (I, II) disease escalation (2 trajectories, 27% and 38% of patients) and (III) stable disease (1 trajectory, 35% of patients). For trajectory prediction from early year data, HMLSs including the stochastic neighbor embedding algorithm (SNEA, as a DRA) as well as locally linear embedding algorithm (LLEA, as a DRA), linked with the new probabilistic neural network classifier (NPNNC, as a classifier), resulted in accuracies of 78.4% and 79.2% respectively, while other HMLSs such as SNEA + Lib_SVM (library for support vector machines) and t_SNE (t-distributed stochastic neighbor embedding) + NPNNC resulted in 76.5% and 76.1% respectively.

CONCLUSIONS

This study moves beyond cross-sectional PD subtyping to clustering of longitudinal disease trajectories. We conclude that combining medical information with SPECT-based radiomics features, and optimal utilization of HMLSs, can identify distinct disease trajectories in PD patients, and enable effective prediction of disease trajectories from early year data.

Feature selection and machine learning methods for optimal identification and prediction of subtypes in Parkinson's disease

OBJECTIVES

The present work focuses on assessment of Parkinson's disease (PD), including both PD subtype identification (unsupervised task) and prediction (supervised task). We specifically investigate optimal feature selection and machine learning algorithms for these tasks.

METHODS

We selected 885 PD subjects as derived from longitudinal datasets (years 0-4; Parkinson's Progressive Marker Initiative), and investigated 981 features including motor, non-motor, and imaging features (SPECT-based radiomics features extracted using our standardized SERA software). Two different hybrid machine learning systems (HMLS) were constructed and applied to the data in order to select optimal combinations in both tasks: (i) identification of subtypes in PD (unsupervised-clustering), and (ii) prediction of these subtypes in year 4 (supervised-classification). From the original data based on years 0 (baseline) and 1, we created new datasets as inputs to the prediction task: (i,ii) CSD0 and CSD01: cross-sectional datasets from year 0 only and both years 0 & 1, respectively; (iii) TD01: timeless dataset from both years 0 & 1. In addition, PD subtype in year 4 was considered as outcome. Finally, high score features were derived via ensemble voting based on their prioritizations from feature selector algorithms (FSAs).

RESULTS

In clustering task, the most optimal combinations (out of 981) were selected by individual FSAs to enable high correlation compared to using all features (arriving at 547). In prediction task, we were able to select optimal combinations, resulting in an accuracy >90% only for timeless dataset (TD01); there, we were able to select the most optimal combination using 77 features, directly selected by FSAs. In both tasks, however, using combination of only high score features from ensemble voting did not enable acceptable performances, showing optimal feature selection via individual FSAs to be more effective.

CONCLUSION

Combining non-imaging information with SPECT-based radiomics features, and optimal utilization of HMLSs, can enable robust identification of subtypes as well as appropriate prediction of these subtypes in PD patients. Moreover, use of timeless dataset, beyond cross-sectional datasets, enabled predictive accuracies over 90%. Overall, we showed that radiomics features extracted from SPECT images are important in clustering as well as prediction of PD subtypes.