| Summary: | Accurate diagnosis of breast cancer has pivotal role in the oncological control strategy. Diagnostic assessment of breast lesion imaging heterogeneity disparities becomes challenging in clinical practice, when dealing with equivocal breast lesions.
Advancements in the radiology field in terms of efficient imaging modalities, as well as innovations in medical image analysis and processing, allow tracking macroscopically, repeatedly, and non-invasively of breast intralesion imaging heterogeneity phenotypes. Specifically, the clinical usage of Magnetic Resonance Imaging (MRI) emerges, as it provides anatomical, functional, and structural information, capturing both spatially and temporally intralesion heterogeneity patterns, through conventional and advanced MR sequences.
Regarding breast MRI, qualitative evaluation of intralesion heterogeneity phenotypes has been endorsed according to the Breast Imaging Reporting and Data System (BI-RADS) Lexicon Classification Form, while quantitative evaluation has been expressed by biomarkers of image signal intensity, shape, and texture characteristics of lesions. Since 1980s, remarkable advances in computer and data analysis algorithms facilitated the extraction of quantifiable image biomarkers/features (i.e., mathematical descriptors) to assess image grey level heterogeneity patterns that possibly indicate underpinning biological processes.
A “new” discipline has emerged by the synergy of quantitative imaging and machine learning, known as radiomics. Radiomics refers to the automatic extraction of a large number of quantitative features from medical images, that are further analyzed and assessed in a high-throughput machine learning framework.
Towards this direction, the current doctorate thesis aims to quantify breast intralesion imaging heterogeneity evaluating intensity-based and texture-based approaches within the framework of radiomics analysis, as an assistive decision support tool to routine multiparametric MRI-based breast cancer diagnosis. Two specific objectives, dealing with challenging concepts regarding exploitation of advanced MRI sequences at 3.0 T, were investigated.
Acknowledging the valuable intrinsic contrast agent free information derived from Diffusion Weighted Imaging (DWI), the first objective of this thesis, was focused on a stand-alone DWI approach, by means of its pixelwise parametric representation (Apparent Diffusion Coefficient, ADC map). DWI quantitative evaluation of breast intralesion heterogeneity for breast cancer diagnosis, is mainly focused on intensity-based descriptors, while texture-based features are barely explored. Thus, the first objective was motivated by the investigation of the additive value of texture-based along with intensity-based features in DWI breast cancer diagnosis.
The patient cohort analyzed, consisted of 78 histologically verified mass like breast lesions (40 benign and 38 malignant). A radiomics-based 2D analysis workflow including steps of intra DW image registration, semi-automated segmentation, feature selection and classification was employed. Specifically, a multi-resolution non-rigid registration scheme, was exploited for mapping the high-b value (b = 900 s/mm2) DW images to corresponding low b-value (b = 0 s/mm2) DW images. Lesion Region of Interest (ROI) segmentation was applied to high b-value (b900) DW images, employing a two-step semi-automated segmentation algorithm, consisting of the Fuzzy C-Means (FCM) followed by an edge/contour-based refinement segmentation method.
Subsequently, ADC maps were generated, while lesion ROI segments were propagated to ADC maps. A total of 27 (11 intensity-based and 16 texture-based) features were extracted. Stepwise feature selection method was employed, while the discriminating ability of features was evaluated with univariate and multivariate Logistic Regression (LR) classification. The classification performance of the diagnostic model was evaluated by means of the Area Under Receiver Operating Characteristic curve (AUC).
Findings demonstrate high classification performance (AUC 0.965), achieved by the feature subset selected by the stepwise regression method, consisting of one intensity-based (25th percentile) of lesion ADC map and one of texture (Grey Level Co-occurrence Matrix -based entropy).
Combining lesion imaging profiles derived from different MRI techniques, such as T2-weighted, DCE, DCE pharmacokinetic models, DWI, ADC map and DKI have been recently suggested for breast cancer differential diagnosis. However, such mpMRI approaches suffer from low clinical translatability, as some of the MR acquisition techniques stated above (pharmacokinetic models and DKI) are rarely met in most institutions worldwide.
Quantitative MR by means of intensity-based and texture-based radiomics analysis in the framework of a commonly utilized clinical MRI acquisition protocol, consisting of DCE and DWI, has not been investigated thoroughly in capturing breast intralesion heterogeneity. In addition, enriched information may be provided incorporating DCE semi-quantitative model-free approaches exploiting pixel-wise empirical indices (parametric maps) derived directly from the time to signal intensity curves (TICs) to deal with the lack of pharmacokinetic models of such MR protocols.
Therefore, the aim of the second specific objective of this doctorate thesis, was to propose a comprehensive intensity-based and texture-based radiomics workflow that enables assessment of breast intralesion heterogeneity, from individual DCE and DWI parametric representations, as well as from their combination (multiparametric MRI, mpMRI representations) in classifying benign and malignant breast lesions.
For this purpose, a patient cohort consisting of 85 histologically verified mass like breast lesions (41 benign and 44 malignant), was employed. The radiomics pipeline that was initially developed for the first specific objective, was utilized, however enriched with the preprocessing step of inter (DCE-DWI) and intra sequence registration and robust machine learning schemes of feature selection (Least Absolute Shrinkage and Selection Operator, LASSO) and three classification schemes (Logistic Regression, LR, Random Forest, RF, Support Vector Machine- Sequential Minimal Optimization, SVM-SMO).
Specifically, in order to capture breast intralesion heterogeneity properties across pixel-wise multiparametric representations, intra- and inter-sequence registration schemes need to be adopted, enabling lesion ROI transfer. Multi-resolution non-rigid (deformable) registration schemes combining three levels of affine, and a final 4th resolution level of a 3rd order B-spline transformation, were employed. In the intra DCE sequence registration process all the DCE time frames (pre- and the 1st, 3rd, 4th, 5th post-contrast) were registered to a single post-contrast slice (the 2nd post-contrast DCE frame, i.e., 2.8 minutes after contrast injection, was considered as the reference image). Similarly, in the inter DCE-DWI sequence registration process the 2nd post-contrast DCE frame was used as the reference image for registering the low b-value (S0) DW image. Finally, in the intra DWI sequence registration process the high b-value (S900) DW image was registered to the previously b0_registered (S0_reg) DW image.
Breast lesions were segmented both manually by an expert and semi-automatically on the2nd post-contrast frame. Manually delineated ROIs serve for the “ground-truth” segments of the analysis. The previously mentioned two-step semi-automated segmentation algorithm, consisting of the Fuzzy C-Means (FCM) followed by an edge/contour-based refinement segmentation method, was employed. The segmentation outcome has been further analyzed in terms of overlap accuracy and reproducibility.
Lesion ROI segments were propagated to DCE semi-quantitative model-free and ADC parametric maps. A total of 27 (11 intensity-based and 16 texture-based) features were extracted from each MR representation, resulting in 135 features for each breast lesion. The robust feature selection method of LASSO was employed, while the discriminating ability of features was evaluated with univariate and multivariate LR, RF and SVM-SMO classification. The classification performance of the diagnostic model was evaluated by means of the AUC.
Findings highlight the dominant role of DCE representations in breast MRI diagnosis, while complementary information is provided with the integration of DWI into the analysis scheme. The mpMRI scheme consisting of seven selected, intensity-based and texture-based entropy and percentile features achieved the highest classification performance across all three (3) classifiers ranging from 0.918 to 0.984. These results prove the feasibility of mpMRI radiomics to contribute to the precision of clinical decision making.
In conclusion, both radiomic approaches, contrast agent free and multiparametric MRI, investigated in the current doctorate thesis, quantify breast intralesion imaging phenotypic heterogeneity, efficiently and accurately. Results suggest that exploiting intensity- based and texture-based radiomics analysis with machine learning, in commonly utilized breast MRI protocols may serve as an assistive decision support tool to MRI-based breast cancer diagnosis.
|
|---|
