PaperlayerView exact revision on arXiv

Source-linked AI summary

Next Generation Radiogenomics Sequencing for Prediction of EGFR and KRAS Mutation Status in NSCLC Patients Using Multimodal Imaging and Machine Learning Approaches

Isaac Shiri, Hassan Maleki, Ghasem Hajianfar, Hamid Abdollahi, Saeed Ashrafinia, Mathieu Hatt, Mehrdad Oveisi, Arman Rahmim

arXiv:1907.02121v1physics.med-phcs.LGphysics.bio-phq-bio.GN

TL;DR

Because existing bio-techniques for assessing EGFR/KRAS mutations are not entirely feasible, the study evaluated whether radiomics from PET and CT images with machine learning could predict mutation status in NSCLC. Radiomics models reached AUCs of 0.82 for EGFR and 0.83 for KRAS, and the authors report greater predictive power than conventional imaging biomarkers.

  • Problem

    Existing bio-techniques for assessing EGFR/KRAS mutation status are not entirely feasible, motivating evaluation of imaging-based prediction in NSCLC patients.

    Source: 8- Model evaluation · p. 38

  • Method

    The authors developed a radiomics framework using low-dose CT, diagnostic CT, and PET images with univariate and multivariate machine-learning algorithms to predict EGFR and KRAS status.

    Source: 8- Model evaluation · p. 38

  • Results

    Radiomics models reached AUCs of 0.82 for EGFR and 0.83 for KRAS, with greater predictive power than conventional imaging biomarkers.

    Source: 8- Model evaluation · pp. 22, 30, 40

  • Takeaways & Limitations

    Radiomic features from different imaging modalities, alone or in combination, could be used for non-invasive prediction of EGFR and KRAS status in NSCLC patients.

    Source: 8- Model evaluation · pp. 38, 40

    The limited dataset size is a concern because analyzing many radiomic features with few imaging data is prone to over-fitting.

    Source: 8- Model evaluation · p. 39

Abstract

from arXiv · show

Aim: In the present work, we aimed to evaluate a comprehensive radiomics framework that enabled prediction of EGFR and KRAS mutation status in NSCLC cancer patients based on PET and CT multi-modalities radiomic features and machine learning (ML) algorithms. Methods: Our study involved 211 NSCLC cancer patient with PET and CTD images. More than twenty thousand radiomic features from different image-feature sets were extracted Feature value was normalized to obtain Z-scores, followed by student t-test students for comparison, high correlated features were eliminated and the False discovery rate (FDR) correction were performed Six feature selection methods and twelve classifiers were used to predict gene status in patient and model evaluation was reported on independent validation sets (68 patients). Results: The best predictive power of conventional PET parameters was achieved by SUVpeak (AUC: 0.69, P-value = 0.0002) and MTV (AUC: 0.55, P-value = 0.0011) for EGFR and KRAS, respectively. Univariate analysis of radiomics features improved prediction power up to AUC: 75 (q-value: 0.003, Short Run Emphasis feature of GLRLM from LOG preprocessed image of PET with sigma value 1.5) and AUC: 0.71 (q-value 0.00005, The Large Dependence Low Gray Level Emphasis from GLDM in LOG preprocessed image of CTD sigma value 5) for EGFR and KRAS, respectively. Furthermore, the machine learning algorithm improved the perdition power up to AUC: 0.82 for EGFR (LOG preprocessed of PET image set with sigma 3 with VT feature selector and SGD classifier) and AUC: 0.83 for KRAS (CT image set with sigma 3.5 with SM feature selector and SGD classifier). Conclusion: We demonstrated that radiomic features extracted from different image-feature sets could be used for EGFR and KRAS mutation status prediction in NSCLC patients, and showed that they have more predictive power than conventional imaging parameters.

1- Data collection

Clinical characteristics of all patients are presented in Table 1.

1-1-Imaging

The study assembled diagnostic CT and PET/CT imaging for 211 NSCLC patients with genomic data. After exclusions for image quality and availability issues, 186 PET and 175 CT and CTD images were segmented.

1-1-Imaging

  • Imaging covered 211 NSCLC patients with genomic data, using diagnostic CT and PET/CT modalities (29-32).PET/CT included low-dose CT for attenuation correction and PET images; scanners were from GE and SIEMENS.Source: 1-1-Imaging · p. 7
  • CTD images were acquired on GE (73%) and SIEMENS (27%) scanners, while PET/CT scans used GE Discovery systems.CTD slice thickness was 0.625–3 mm (median: 1.5 mm); PET acquisition used 18F-FDG doses of 138.90–572.25 MBq (mean: 309.26 MBq) and uptake times of 23.08–128.90 min (mean: 66.58 minutes).Source: 1-1-Imaging · p. 7
  • Image noise, artifacts, missing sequences, and mis-segmentation led to exclusions, leaving 186 PET and 175 low-dose CT and CTD images segmented.Source: 1-1-Imaging · p. 7

1-2-Genomics

Tumor samples were tested for EGFR and KRAS mutations using SNaPshot single-base extension after multiplex PCR.

1-2-Genomics

  • SNaPshot testing assessed EGFR exons 18–21 and KRAS exon 2 positions 12 and 13 for missense mutations.Tumors were excised, cut into 3–5 mm slices, and frozen for 30 minutes before testing.Source: 1-2-Genomics · p. 7

2- Image pre-processing

Images were standardized and transformed into multiple feature sets before radiomic feature extraction.

2- Image pre-processing

  • Images were resampled to isotropic voxel spacings (CTD 1×1×1, CT 1×1×3, PET 3×3×3), then processed with LOG, wavelet, and 64-bin discretization filters.LOG sigma values ranged from 0.5 to 5 in 0.5 steps; wavelet filtering produced eight three-dimensional high/low-pass decompositions. Preprocessing was applied to intensity, histogram, and textural features.Source: 2- Image pre-processing · p. 7

3- Image segmentation

PET lesions were manually delineated, while CT images were automatically region-grown and then edited and verified by a radiologist.

3- Image segmentation

  • PET lesions were manually delineated in OSIRIX®, while CT images underwent automatic region growing in 3D-Slicer, followed by radiologist editing and verification; the study provided 186 PET and 175 low-dose CT and CTD segmentation regions.Source: 3- Image segmentation · p. 7

4- Feature extraction

The study extracted radiomic features across seven feature classes alongside conventional PET biomarkers, using PyRadiomics and IBSI-compliant feature definitions.

4- Feature extraction

  • Features spanned 19 FOS, 16 shape, 23 GLCM, 16 GLRLM, 16 GLSZM, 5 NGTDM, and 14 GLDM measures.Feature extraction used the open-source PyRadiomics library.Source: 4- Feature extraction · p. 7
  • PyRadiomics features comply with IBSI definitions, supporting harmonization and reproducibility of the calculated radiomic features.The passage states this facilitates reproducibility of the study.Source: 4- Feature extraction · p. 7
  • Conventional PET biomarkers included MTV, SUVmax, SUVpeak, SULmax, and SULpeak.SUL values represent normalization to lean body mass.Source: 4- Feature extraction · p. 7

5- Univariate analysis

The univariate analysis standardized features, tested differences, removed highly correlated features, applied FDR correction, and evaluated conventional clinical PET parameters.

5- Univariate analysis

  • Features were converted to Z-scores, compared using Student’s t-tests with p < 0.05 as the significance threshold, filtered for high Spearman correlation, and assessed with FDR correction and reported q-values.Statistical analyses used R 3.5.1 with the pROC and stats packages.Source: 5- Univariate analysis · p. 9
  • The analysis also evaluated SUVmax, SUVpeak, SULmax, SULpeak, and MTV as conventional clinical PET parameters.Source: 5- Univariate analysis · p. 9

6- Feature selection

The framework compared six feature-selection methods: three filter-based selectors, one embedded selector, and two combinations of variance thresholding with other selectors.

Feature selection

  • Six selectors were implemented: Select K Best (SKB), Variance Threshold (VT), Select Percentile (SP), Select from Model (SM), VT-SM, and VT-KB.They comprised three filter-based methods (SKB, VT, SP), one embedded method (SM), and two combined methods (VT-SM and VT-KB).Source: 6- Feature selection · p. 9
  • Select Percentile retains a user-specified highest-scoring percentage of features using common univariate statistical tests.Source: 6- Feature selection · p. 9
  • Select from Model removes features whose fitted coefficients or feature-importance values fall below a threshold, which may be specified numerically or by heuristics such as mean or median.Source: 6- Feature selection · p. 9

7- Classifier

The study implemented and compared 12 classifiers spanning linear, naive Bayes, nearest-neighbor, tree, discriminant, support-vector, neural-network, and ensemble methods.

7- Classifier

  • The 12 classifiers included logistic regression, stochastic gradient descent, naive Bayes, Gaussian naive Bayes, and k-nearest neighbors.Source: 7- Classifier · p. 9
  • The classifier set also included C5.0 decision trees, quadratic discriminant analysis, support vector classifiers, and multi-layer perceptrons.Source: 7- Classifier · p. 9
  • Ensemble methods comprised adaptive boosting, bagging, and random forest.Source: 7- Classifier · p. 9

8- Model evaluation

Radiomic machine-learning models predicted EGFR and KRAS mutation status in NSCLC patients, with the best reported models outperforming conventional PET parameters. Performance varied across radiomic features and model combinations, and models were evaluated on independent validation sets.

Model evaluation

  • Models were tuned with 10-fold cross-validation repeated 20 times and evaluated on independent validation sets; AUC summarized predictive performance.The evaluation paragraph reports 65 validation patients, while the later limitations paragraph reports 63.Source: 8- Model evaluation · pp. 10, 39

Univariate radiomics results

  • About 90 EGFR-associated features had mean AUC 0.67±0.05, while 14 KRAS-associated features had mean AUC 0.67±0.02 after correlation filtering and FDR correction.The top EGFR feature was Short Run Emphasis of GLRLM from PET_LOG sigma 1.5 (q-value 0.003; AUC: 75); the top KRAS feature was Large Dependence Low Gray Level Emphasis of GLDM from CTD_LOG sigma 5 (AUC 0.71; q-value 0.00005).Source: 8- Model evaluation · p. 11

Shape features of EGFR and KRAS

  • Shape features were not statistically significant: Surface Volume Ratio from PET had AUC 0.60 for EGFR, and CT Flatness had AUC 0.67 for KRAS.Their q-values were 0.28 and 0.11, respectively.Source: 8- Model evaluation · p. 20

Conventional PET models

  • AUC 0.69 for SUVpeak was the best conventional PET result for EGFR, while MTV had the best KRAS result at AUC 0.55.SUVpeak had P-value 0.0002 for EGFR; MTV had P-value 0.0011 for KRAS.Source: 8- Model evaluation · p. 22

Multivariate machine learning radiomic models

Feature selection and classifier results

  • Across feature-selector and classifier combinations, the highest reported mean AUCs were 0.73±0.074 for EGFR selection and 0.76±0.024 for EGFR classification, both using PET_W_LHH.For KRAS, the highest selector mean was 0.66±0.093 with SM and PET_LOG_4.5S, while the highest classifier mean was 0.79±0.021 with SVM and CT_LOG_3.5S.Source: 8- Model evaluation · pp. 24, 30

Related papers

Loading 1907.02121v1…