Standardised lesion segmentation for imaging biomarker quantitation: a consensus recommendation from ESR and EORTC.

BACKGROUND: Lesion/tissue segmentation on digital medical images enables biomarker extraction, image-guided therapy delivery, treatment response measurement, and training/validation for developing artificial intelligence algorithms and workflows. To ensure data reproducibility, criteria for standardised segmentation are critical but currently unavailable. METHODS: A modified Delphi process initiated by the European Imaging Biomarker Alliance (EIBALL) of the European Society of Radiology (ESR) and the European Organisation for Research and Treatment of Cancer (EORTC) Imaging Group was undertaken. Three multidisciplinary task forces addressed modality and image acquisition, segmentation methodology itself, and standards and logistics. Devised survey questions were fed via a facilitator to expert participants. The 58 respondents to Round 1 were invited to participate in Rounds 2-4. Subsequent rounds were informed by responses of previous rounds. RESULTS/CONCLUSIONS: Items with ≥ 75% consensus are considered a recommendation. These include system performance certification, thresholds for image signal-to-noise, contrast-to-noise and tumour-to-background ratios, spatial resolution, and artefact levels. Direct, iterative, and machine or deep learning reconstruction methods, use of a mixture of CE marked and verified research tools were agreed and use of specified reference standards and validation processes considered essential. Operator training and refreshment were considered mandatory for clinical trials and clinical research. Items with a 60-74% agreement require reporting (site-specific accreditation for clinical research, minimal pixel number within lesion segmented, use of post-reconstruction algorithms, operator training refreshment for clinical practice). Items with ≤ 60% agreement are outside current recommendations for segmentation (frequency of system performance tests, use of only CE-marked tools, board certification of operators, frequency of operator refresher training). Recommendations by anatomical area are also specified.

Integrating segmentation information into CNN for breast cancer diagnosis of mammographic masses.

Background and ObjectivesSegmentation of mammographic lesions has been proven to be a valuable source of information, as it can assist in both extracting shape-related features and providing accurate localization of the lesion. In this work, a methodology is proposed for integrating mammographic mass segmentation information into a convolutional neural network (CNN), aiming to improve the diagnosis of breast cancer in mammograms. MethodsThe proposed methodology involves modification of each convolutional layer of a CNN, so that information of not only the input image but also the corresponding segmentation map is considered. Furthermore, a new loss function is introduced, which adds an extra term to the standard cross-entropy, aiming to steer the attention of the network to the mass region, penalizing strong feature activations based on their location. The segmentation maps are acquired either from the provided ground-truth or from an automatic segmentation stage. ResultsPerformance evaluation in diagnosis is conducted on two mammographic mass datasets, namely DDSM-400 and CBIS-DDSM, with differences in quality of the corresponding ground-truth segmentation maps. The proposed method achieves diagnosis performance of 0.898 and 0.862 in terms AUC when using ground-truth segmentation maps and a maximum of 0.880 and 0.860 when a U-Net-based automatic segmentation stage is employed, for DDSM-400 and CBIS-DDSM, respectively. ConclusionsThe experimental results demonstrate that integrating segmentation information into a CNN leads to improved performance in breast cancer diagnosis of mammographic masses.

Multi-parametric MRI lesion heterogeneity biomarkers for breast cancer diagnosis.

PURPOSE: To identify intra-lesion imaging heterogeneity biomarkers in multi-parametric Magnetic Resonance Imaging (mpMRI) for breast lesion diagnosis. METHODS: Dynamic Contrast Enhanced (DCE) and Diffusion Weighted Imaging (DWI) of 73 female patients, with 85 histologically verified breast lesions were acquired. Non-rigid multi-resolution registration was utilized to spatially align sequences. Four (4) DCE (2nd post-contrast frame, Initial-Enhancement, Post-Initial-Enhancement and Signal-Enhancement-Ratio) and one (1) DWI (Apparent-Diffusion-Coefficient) representations were analyzed, considering a representative lesion slice. 11 1st-order-statistics and 16 texture features (Gray-Level-Co-occurrence-Matrix (GLCM) and Gray-Level-Run-Length-Matrix (GLRLM) based) were derived from lesion segments, provided by Fuzzy C-Means segmentation, across the 5 representations, resulting in 135 features. Least-Absolute-Shrinkage and Selection-Operator (LASSO) regression was utilized to select optimal feature subsets, subsequently fed into 3 classification schemes: Logistic-Regression (LR), Random-Forest (RF), Support-Vector-Machine-Sequential-Minimal-Optimization (SVM-SMO), assessed with Receiver-Operating-Characteristic (ROC) analysis. RESULTS: LASSO regression resulted in 7, 6 and 7 features subsets from DCE, DWI and mpMRI, respectively. Best classification performance was obtained by the RF multi-parametric scheme (Area-Under-ROC-Curve, (AUC) ± Standard-Error (SE), AUC ± SE = 0.984 ± 0.025), as compared to DCE (AUC ± SE = 0.961 ± 0.030) and DWI (AUC ± SE = 0.938 ± 0.032) and statistically significantly higher as compared to DWI. The selected mpMRI feature subset highlights the significance of entropy (1st-order-statistics and 2nd-order-statistics (GLCM)) and percentile features extracted from 2nd post-contrast frame, PIE, SER maps and ADC map. CONCLUSION: Capturing breast intra-lesion heterogeneity, across mpMRI lesion segments with 1st-order-statistics and texture features (GLCM and GLRLM based), offers a valuable diagnostic tool for breast cancer.

Quantitative assessment of microcalcification cluster image quality in digital breast tomosynthesis, 2-dimensional and synthetic mammography.

Quantitative assessment of microcalcification (MC) cluster image quality is presented, in terms of cluster signal-difference-to-noise ratio (SDNR) intercomparison among digital breast tomosynthesis (DBT) and 2-dimensional (2D) and synthetic-2-dimensional (s2D) mammography. A phantom that provides realistic appearance of MC clusters located in uniform and nonuniform background was imaged in 2D and DBT, considering various scattering conditions. MC cluster SDNR differentiation is investigated with respect to MC particle size (uniform background) and surrounding parenchyma density (nonuniform background). An accurate MC cluster segmentation method was used to delineate individual MC particles and estimate MC cluster SDNR. Analysis of the uniform part of the phantom indicated higher performance of DBT and 2D over s2D for the smallest cluster size (106-177 µm), no difference among mammographic modes for the largest MC cluster (224-354 µm), and enhanced role of 2D for decreasing cluster size and increasing scattering. Analysis of the nonuniform part of the phantom indicated DBT performed better than 2D and s2D in case of dense parenchyma pattern, while 2D and s2D did not differ across parenchyma density patterns and scattering conditions. The presented MC cluster SDNR analysis was capable of revealing subtle differences among mammographic modes and suggests a methodology for clinical image quality assessment. Graphical abstract.

Deep Learning for Breast Cancer Diagnosis from Mammograms-A Comparative Study.

Deep convolutional neural networks (CNNs) are investigated in the context of computer-aided diagnosis (CADx) of breast cancer. State-of-the-art CNNs are trained and evaluated on two mammographic datasets, consisting of ROIs depicting benign or malignant mass lesions. The performance evaluation of each examined network is addressed in two training scenarios: the first involves initializing the network with pre-trained weights, while for the second the networks are initialized in a random fashion. Extensive experimental results show the superior performance achieved in the case of fine-tuning a pretrained network compared to training from scratch.

Quantitative visually lossless compression ratio determination of JPEG2000 in digitized mammograms.

The current study presents a quantitative approach towards visually lossless compression ratio (CR) threshold determination of JPEG2000 in digitized mammograms. This is achieved by identifying quantitative image quality metrics that reflect radiologists' visual perception in distinguishing between original and wavelet-compressed mammographic regions of interest containing microcalcification clusters (MCs) and normal parenchyma, originating from 68 images from the Digital Database for Screening Mammography. Specifically, image quality of wavelet-compressed mammograms (CRs, 10:1, 25:1, 40:1, 70:1, 100:1) is evaluated quantitatively by means of eight image quality metrics of different computational principles and qualitatively by three radiologists employing a five-point rating scale. The accuracy of the objective metrics is investigated in terms of (1) their correlation (r) with qualitative assessment and (2) ROC analysis (A z index), employing pooled radiologists' rating scores as ground truth. The quantitative metrics mean square error, mean absolute error, peak signal-to-noise ratio, and structural similarity demonstrated strong correlation with pooled radiologists' ratings (r, 0.825, 0.823, -0.825, and -0.826, respectively) and the highest area under ROC curve (A z , 0.922, 0.920, 0.922, and 0.922, respectively). For each quantitative metric, the highest accuracy values of corresponding ROC curves were used to define metric cut-off values. The metrics cut-off values were subsequently used to suggest a visually lossless CR threshold, estimated to be between 25:1 and 40:1 for the dataset analyzed. Results indicate the potential of the quantitative metrics approach in predicting visually lossless CRs in case of MCs in mammography.

Size-adapted microcalcification segmentation in mammography utilizing scale-space signatures.

The purpose of this study is size-adapted segmentation of individual microcalcifications in mammography, based on microcalcification scale-space signature estimation, enabling robust scale selection for initialization of multiscale active contours. Segmentation accuracy was evaluated by the area overlap measure, by comparing the proposed method and two recently proposed ones to expert manual delineations. The method achieved area overlap of 0.61+/-0.15 outperforming statistically (p<0.001) the other two methods (0.53+/-0.18, 0.42+/-0.16). Only the proposed method performed equally for both small (< 460 microm) and large (>/= 460 microm) microcalcifications. Results indicate an accurate method, which could be utilized in computer-aided diagnosis schemes of microcalcification clusters.

Atlas-based segmentation of degenerated lumbar intervertebral discs from MR images of the spine.

Intervertebral disc degeneration is an age-associated condition related to chronic back pain, while its consequences are responsible for over 90 % of spine surgical procedures. In clinical practice, MRI is the modality of reference for diagnosing disc degeneration. In this study, we worked toward 2-D semiautomatic segmentation of both normal and degenerated lumbar intervertebral discs from T2-weighted midsagittal MR images of the spine. This task is challenged by partial volume effects and overlapping gray-level values between neighboring tissue classes. To overcome these problems three variations of atlas-based segmentation using a probabilistic atlas of the intervertebral disc were developed and their accuracies were quantitatively evaluated against manually segmented data. The best overall performance, when considering the tradeoff between segmentation accuracy and time efficiency, was accomplished by the atlas-robust-fuzzy c-means approach, which combines prior anatomical knowledge by means of a rigidly registered probabilistic disc atlas with fuzzy clustering techniques incorporating smoothness constraints. The dice similarity indexes of this method were 91.6 % for normal and 87.2 % for degenerated discs. Research in progress utilizes the proposed approach as part of a computer-aided diagnosis system for quantification and characterization of disc degeneration severity. Moreover, this approach could be exploited in computer-assisted spine surgery.

B-spline active rays segmentation of microcalcifications in mammography.

Accurate segmentation of microcalcifications in mammography is crucial for the quantification of morphologic properties by features incorporated in computer-aided diagnosis schemes. A novel segmentation method is proposed implementing active rays (polar-transformed active contours) on B-spline wavelet representation to identify microcalcification contour point estimates in a coarse-to-fine strategy at two levels of analysis. An iterative region growing method is used to delineate the final microcalcification contour curve, with pixel aggregation constrained by the microcalcification contour point estimates. A radial gradient-based method was also implemented for comparative purposes. The methods were tested on a dataset consisting of 149 mainly pleomorphic microcalcification clusters originating from 130 mammograms of the DDSM database. Segmentation accuracy of both methods was evaluated by three radiologists, based on a five-point rating scale. The radiologists' average accuracy ratings were 3.96 +/- 0.77, 3.97 +/- 0.80, and 3.83 +/- 0.89 for the proposed method, and 2.91 +/- 0.86, 2.10 +/- 0.94, and 2.56 +/- 0.76 for the radial gradient-based method, respectively, while the differences in accuracy ratings between the two segmentation methods were statistically significant (Wilcoxon signed-ranks test, p < 0.05). The effect of the two segmentation methods in the classification of benign from malignant microcalcification clusters was also investigated. A least square minimum distance classifier was employed based on cluster features reflecting three morphological properties of individual microcalcifications (area, length, and relative contrast). Classification performance was evaluated by means of the area under ROC curve (Az). The area and length morphologic features demonstrated a statistically significant (Mann-Whitney U-test, p < 0.05) higher patient-based classification performance when extracted from microcalcifications segmented by the proposed method (0.82 +/- 0.06 and 0.86 +/- .05, respectively), as compared to segmentation by the radial gradient-based method (0.71 +/- 0.08 and 0.75 +/- 0.08). The proposed method demonstrates improved segmentation accuracy, fulfilling human visual criteria, and enhances the ability of morphologic features to characterize microcalcification clusters.

Breast cancer diagnosis: analyzing texture of tissue surrounding microcalcifications.

The current study investigates texture properties of the tissue surrounding microcalcification (MC) clusters on mammograms for breast cancer diagnosis. The case sample analyzed consists of 85 dense mammographic images, originating from the Digital Database for Screening Mammography. Mammograms analyzed contain 100 subtle MC clusters (46 benign and 54 malignant). The tissue surrounding MCs is defined on original and wavelet decomposed images, based on a redundant discrete wavelet transform. Gray-level texture and wavelet coefficient texture features at three decomposition levels are extracted from surrounding tissue regions of interest (ST-ROIs). Specifically, gray-level first-order statistics, gray-level cooccurrence matrices features, and Laws' texture energy measures are extracted from original image ST-ROIs. Wavelet coefficient first-order statistics and wavelet coefficient cooccurrence matrices features are extracted from subimages ST-ROIs. The ability of each feature set in differentiating malignant from benign tissue is investigated using a probabilistic neural network. Classification outputs of most discriminating feature sets are combined using a majority voting rule. The proposed combined scheme achieved an area under receiver operating characteristic curve ( A(z)) of 0.989. Results suggest that MCs' ST texture analysis can contribute to computer-aided diagnosis of breast cancer.

Texture classification-based segmentation of lung affected by interstitial pneumonia in high-resolution CT.

Accurate and automated lung field (LF) segmentation in high-resolution computed tomography (HRCT) is highly challenged by the presence of pathologies affecting lung borders, also affecting the performance of computer-aided diagnosis (CAD) schemes. In this work, a two-dimensional LF segmentation algorithm adapted to interstitial pneumonia (IP) patterns is presented. The algorithm employs k-means clustering followed by a filling operation to obtain an initial LF order estimate. The final LF border is obtained by an iterative support vector machine neighborhood labeling of border pixels based on gray level and wavelet coefficient statistics features. A second feature set based on gray level averaging and gradient features was also investigated to evaluate its effect on segmentation performance of the proposed method. The proposed method is evaluated on a dataset of 22 HRCT cases spanning a range of IP patterns such as ground glass, reticular, and honeycombing. The accuracy of the method is assessed using area overlap and shape differentiation metrics (d(mean), d(rms), and d(max)), by comparing automatically derived lung borders to manually traced ones, and further compared to a gray level thresholding-based (GLT-based) method. Accuracy of the methods evaluated is also compared to interobserver variability. The proposed method incorporating gray level and wavelet coefficient statistics demonstrated the highest segmentation accuracy, averaged over left and right LFs (overlap=0.954, d(mean)=1.080 mm, d(rms)=1.407 mm, and d(max)=4.944 mm), which is statistically significant (two-tailed student's t test for paired data, p<0.0083) with respect to all metrics considered as compared to the proposed method incorporating gray level averaging and gradient features (overlap=0.918, d(mean)=2.354 mm, d(rms)=3.711 mm, and d(max)=14.412 mm) and the GLT-based method (overlap=0.897, d(mean)=3.618 mm, d(rms)=5.007 mm, and d(max)=16.893 mm). The performance of the three segmentation methods, although decreased as IP pattern severity level (mild, moderate, and severe) was increased, did not demonstrate statistically significant difference (two-tailed student's t test for unpaired data, p>0.0167 for all metrics considered). Finally, the accuracy of the proposed method, based on gray level and wavelet coefficient statistics ranges within interobserver variability. The proposed segmentation method could be used as an initial stage of a CAD scheme for IP patterns.

A digital equalisation technique improving visualisation of dense mammary gland and breast periphery in mammography.

INTRODUCTION: In mammographic imaging, use of high contrast screen-film combinations results in under-exposed and over-exposed film areas corresponding to dense mammary gland and breast periphery (BP), respectively, characterised by degraded image contrast. A digital equalisation technique was designed and developed in order to deal with the problem of poor visualisation of these regions. METHODS AND MATERIAL: The technique is based on the film-digitiser characteristic curve and a layer model of the breast region, as depicted on a mammogram. It remaps each layer grey level (GL) values by a correction factor that accounts for thickness variation in BP and the presence of dense fibroglandular tissues at the mammary gland. The major steps of the technique are segmentation, to isolate the breast region from mammogram background, and adaptive layer GL remapping. RESULTS: The performance of the technique was initially evaluated on a sample of 60 mammograms. Comparative evaluation between the initial and processed images was performed on the basis of nine anatomical features situated at dense mammary gland and BP. The mammographic images resulting from application of the proposed technique are GL equalised and the visualisation improvement of all anatomical features was found to be statistically significant (P<0.05) or highly significant (P<0.0001). The proposed technique was also compared with contrast-limited adaptive histogram equalisation (CLAHE) and found to be more effective in the visualisation of all anatomical features examined, for both dense breast (DB) and BP. DISCUSSION AND CONCLUSION: Application of the proposed technique results in improved visualisation of both dense mammary gland and BP regions. The proposed technique is independent of breast size, breast symmetry and mammographic view. The technique contributes to breast dose minimisation by eliminating the need for a second acquisition.