Classification of Benign and Malignant DCE-MRI Breast Tumors by Analyzing the Most Suspect Region

Transcription

1 Classification of Benign and Malignant DCE-MRI Breast Tuors by Analyzing the Most Suspect Region Sylvia Glaßer 1, Uli Nieann 1, Uta Prei 2, Bernhard Prei 1, Myra Spiliopoulou 3 1 Departent for Siulation and Graphics, OvG-University Magdeburg 2 Departent for Radiology, Municipal Hospital Magdeburg 3 Knowledge Manageent and Discovery Lab (KMD), OvG-University Magdeburg glasser@isg.cs.uni-agdeburg.de Abstract. Classification of breast tuors solely based on dynaic contrast enhanced agnetic resonance data is a challenge in clinical research. In this paper, we analyze how the ost suspect region as group of siilarly perfused and spatially connected voxels of a breast tuor contributes to distinguishing between benign and alignant tuors. We use three density-based clustering algoriths to partition a tuor in regions and depict the ost supect one, as delivered by the ost stable clustering algorith. We use the properties of this region for each tuor as input to a classifier. Our preliinary results show that the classifier separates between benign and alignant tuors, and returns predictive attributes that are intuitive to the expert. 1 Introduction Dynaic contrast enhanced agnetic resonance iaging (DCE-MRI) allows for perfusion characterization of breast tuors. DCE-MRI has high sensitivity but oderate specificity. So, it reains suppleental to conventional X-ray aography, and is frequently used to confir the alignancy or benignity of lesions [1]. Malignant breast tuors often lead to neo-angiogenesis with increased tissue pereability and increased nuber of supporting blood vessels; this is usually reflected in a rapid contrast agent washin and/or washout. Hence, it is typical to define a region of interest (ROI) and copute the ROI s average relative enhanceent (RE) over tie - the RE curve - for it. Fro the early RE and the curve s shape, the radiologist assesses the contrast agent washin and washout. Since a breast tuor is as alignant as its ost alignant part, the RE curve of this ROI is used to deterine the tuor s alignancy. In this study, we partition a ROI into regions that are hoogeneous with respect to the RE curves of their voxels, and identify region features that contribute to predict alignancy. The RE curves of the individual voxels are noisy by nature, so a ajor challenge lays in grouping the to hoogeneous and spatially contiguous regions.

2 2 Glaßer et al. Glaßer et al. propose a region erging ethod to this purpose [3], which is used in [4] to study the role of tuor heterogeneity in predicting alignancy. Siilar to our work is the study of Chen et al. [2], who perfor clustering with fuzzy c-eans and extract the ost characteristic RE curve for the separation between benign and alignant tuors. However, our approach cobines the identification of the ost suspect region per tuor with the identification of predictive tuor characteristics that hold for ultiple tuors. In this paper, we study a set of 68 breast tuors. For each tuor, we apply ultiple density-based clustering algoriths and identify the ost suspect region. We extract the properties of this region and show that these properties contribute to distinguishing between benign and alignant tuors. To do so, we perfor classification over all tuors, whereby we blend out obvious predictors like tuor size. Our focus relies on clinical research on DCE-MRI tuor enhanceent kinetics. In clinical practice, rather, the cobination of different iage odalities (e.g. X-ray and MRI) with patient-specific attributes (e.g. genetic risk factors) is indispensable for a coplete diagnosis. 2 Material and Methods In this section we describe our edical iage data containing the breast tuors, followed by our classification approach based on a tuor s ost suspect region. 2.1 Tuor Data Ourdatasetcoprises50patientswith68breasttuors.31tuorsprovedtobe benign and 37 alignant (confiration was carried out via histopathologic evaluation or by follow up studies after six to nine onths). We included only lesions that have been detected in MRI. The data sets were acquired with a 1.0 T open MR scanner and exhibit the paraeters: in-plane resolution , atrix ,nuber of slices 100,slice gap = 1.5, nuber ofacquisitions = 5 6 and total acquisition tie 400sec. During and iediately after the bolus injection of contrast agent one precontrast and four to five post-contrast iages were acquired per series. Since DCE-MRI data exhibit otion artifacts ainly due to thorax expansion through breathing and patient s oveent, otion correction was carried out with MeVisLab ( eploying the elastic registration developed by Rueckert et al. [5]. Next, the relative enhanceent (RE) of a tuor, i.e. the percent aged signal intensity increase, is calculated [1] with RE = (SI c SI)/SI 100. Here, SI is the pre-contrast and SI c is the post-contrast signal intensity. Each breast tuor was segented by an experienced radiologist. The segentation coprises only voxels exhibiting at least 50% RE at the first tie step after the early post contrast phase. 2.2 Methods Our approach consists of two steps: the extraction of the ost suspect region for each tuor and the classification process over all tuors (see Fig. 1).

3 Region-Based Classification of Benign and Malignant Breast Tuors 3 1. Get Most Suspect Region 50 DCE- MRI Data Sets Tuor Segentation Extract Features For Dataset Enrichent 68 Breast Tuors Extraction of Perfusion Paraeters Extraction of 3TP Classes 2. Classification 10-Fold-Cross-Validation 68 Most Suspect Regions Training Set Test Set Densitybased Clustering Learn Classifier Selection of Most Suspect Region ζ Classification Region Foration Distinguish Tuors With Decision Tree Fig. 1. Scheatic overview of the presented approach. First, we deterine each tuor s ost suspect region. Second, we learn a classifier to predict alignancy. Step 1: Extraction of the Most Suspect Region of Each Tuor. We extract the ost suspect region by deterining descriptive perfusion paraeters, three-tie-point classes and applying density-based clustering. The RE plotted over tie yields RE curves that allow for the extraction of the descriptive perfusion paraeters (see Fig. 2(a)): washin (the steepness of the ascending curve), washout (the steepness of the descending curve), peak enhanceent (the axiu RE value), integral (the area under the curve) and tie to peak (the tie when peak enhanceent occurs), which are substitutes for physiological paraeters like tuor perfusion and vessel pereability. Since peak enhanceent and integral strongly correlate, we exclude peak enhanceent fro the subsequent analysis. RE [%] RE [%] Peak Enhanceent (a) Tie to Peak Washin Integral Relative Enhanceent Curve Washout tie (b) fast RE(t ) > 100% 2 9 Increasing Curve 8 +/- 10% Plateau Curve 7 Washout Curve 6 noral 5 50% < RE(t 2) < 100% 4 3TP classes slow RE(t 3 ) < 50% tie t 1 t 2 t 3 Fig.2. In (a) a RE curve and its descriptive perfusion paraeters are depicted. In (b), the 3TP classes based on RE at t 1, t 2, and t 3 are presented. The three-tie-point (3TP) ethod presented by Degani et al. [6] allows for an autoatic RE curve classification based on three well chosen tie points: t 1, the first point in tie before the contrast agent injection, t 2, 2in after t 1 and t 3, 4in after t 2. With the 3TP ethod, a RE change in the interval ±10% in the tie between t 2 and t 3 will be interpreted as plateau, whereas RE changes higher than 10% and lower than 10% [6] are classified as increasing curve and washout curve, respectively. Since our study contains 5-6 tie steps due to different scanning paraeters, we assign the third tie step to t 2 and the last tie step to t 3. The analysis of the initial contrast agent accuulation, i.e. the RE value at t 2, which is classified into slow, noral and fast in cobination

4 4 Glaßer et al. with the three curve shapes yields nine curve types (see Fig. 2(b)). We copare the results of our clustering algoriths with the 3TP classes. For the evaluation of the tuor enhanceent, voxels with siilar RE values should be grouped into regions. To account for irregular and heterogeneous tuor parts, we adapted and applied the following density-based clustering algoriths to all data sets of our study: Density-based Spatial Clustering of Applications with ise (DBSCAN) [7], Density-Connected Subspace Clustering (SUB- CLU)[8], and Ordering Points to Identify the Clustering Structure (OPTICS)[9]. The algoriths separate objects into clusters based on estiated density distributions. They yield clusters with arbitrary shapes, which is iportant for the underlying edical iage data. Objects that do not feature siilar objects (i.e. objects with siilar paraeters) in a given neighbourhood are arked as outliers. That s a further advantage, since outliers ay be caused by a issing inter-voxel-correspondence over tie due to otion artifacts. Next, we eploy each voxel s perfusion paraeters and its relative position in the data set as observations. We apply DBSCAN, SUBCLU and OPTICS with the following paraeters: the nuber of iniu points inp ts for a cluster is set to 4, 6, and 8. The ǫ-value that deterines the size of the neighbourhood depends on the clustering. For DBSCAN and each inpts value, ǫ was autoatically deterined as suggested in [7]. For SUBCLU, ǫ was extracted fro the k-distancesgraph[7]thatapsthedistanceofanobjecttoitsknextneighbours. OurautoaticestiationisdepictedinFig.3.Weapplythisapproachtoallfour perfusion paraeter sets and assign ǫ to the ean of the four estiated values. For OPTICS, we epirically set ǫ to 0.5 and With inpts {4,6,8}, we get three configurations for DBSCAN and SUBCLU and six configurations for OPTICS yielding 12 clustering results per data set. Spatially connected clusters are aintained by a connected coponent analysis. Fig. 3. Deterination of ǫ based on the k-distances graph for a given inpts value. The graph aps the distance of an object to its k next neighbours (with k = inpts). A well suited ǫ can be detected at a position with increased slope. It is autoatically deterined by choosing the point with the biggest distance perpendicular to a line g connecting the first and the last point of the graph. ɛ k-distance Graph g objects To select the ost suspect region, we choose the clustering with the least outliers. Next, we reject all regions that contain less than three voxels. Fro the reaining regions, we choose the biggest region with an average RE curve of 3TP class 7. If no such region exists, we search for the 3TP class 9, 8, 4, 6, 5, 1, 3, 2 in that order. Although this is a user-defined ranking, we establish this epirical ranking based on definitions of the ost alignant tuor enhanceent kinetics: a present washout in cobination with a strong contrast washin (see Fig. 2(b)). Step 2: Data Enrichent and Classifier Learning. In the second part, we cobine the extracted data to learn a classifier ζ over our 68 breast lesions.

5 Region-Based Classification of Benign and Malignant Breast Tuors 5 Data enrichent is carried out by including the following attributes of the tuor and its ost suspect region: tuor size (in 3 ), nuber of tuor voxels, percent aged region size, nuber of region voxels, the siilarity easures Purity (P), Jaccard index (J) and the F1 score (F 1 ) based on the coparison of our clustering and the 3TP-based division (we extracted and include values per tuor, per region and per outlier cluster), the region s ean perfusion paraeters, the nuber of region voxels of each of the 3TP classes, the ost proinent 3TP class (the 3TP class to which the ajority of region voxels belong), the region s ean RE curve, the RE curve s 3TP class and the patient s age. For our approach, we use the J4.8 classification algorith of the Waikato Environent for Knowledge Analysis (Weka) library - a Java software library that encopasses algoriths for data analysis and predictive odelling [10]. J4.8 is based on the C4. 5 decision tree classification [11]. It perfors 10-fold cross validation and requires at least two instances (two tuors) for each tree leaf. We worked iteratively, reducing the features under consideration and aiing to avoid features that are obviously predictive (such as tuor size), so that the predictive power of other features is highlighted. As a result, we cae up with the classifier ζ that was learned on 18 of the original ca. 40 features. b 14 0 RE(t 3) > 102% 2 1 age > 61 b 5 0 J Outlier > 0.05 Suary - Correctly classified instances: 46 / 68 (67.65%) - Incorrectly classified instances: 22 / 68 (32,35%) - True Positive (TP) and False Positive (FP) rate: - benign: TP = 0.548, FP = alignant: TP = 0.784, FP = J Tuor > Majority of 3TP clas case 7 case 9 case 6 case 8 default #(3TP ) > 56 6 b Label (b=benign; =alignant) with nuber of correct ( ) and incorrect ( ) classified tuors of the leaf #(3TP ) > 4 9 b 3 0 Washout > Fig.4. Learned decision tree: the attributes at the upper part of the tree are the ost iportant ones. 3 Results We learned a decision tree (see Fig. 4) that eploys eight of the 18 features and classifies 46 of the 68 lesions correctly. Features closer to the root of the tree are ore iportant than those at lower levels, because the forer help in splitting a larger set of tuors. We can see that the ost iportant attributes are the heterogeneity of the tuor, as identified by the clustering algorith(represented by J Tuor and J Outlier ) and the age of the patient. The ost proinent 3TP class, the nuber of voxels in the 3TP class 9 (#(3TP 9 )) and 6 (#(3TP 6 )), the contrast agent washin (RE(t 3 )), and washout are also iportant.

6 6 Glaßer et al. 4 Discussion We presented a new ethod that cobines within-tuor clustering and tuor classification to predict tuor alignancy, and we reported on our preliinary results on a data set with 68 breast tuors. Our first results indicate that the identification of the ost suspect region with clustering and the exploitation of this region s features in classification are proising steps in tuor separation. The low sensitivity of our results ust be attributed to the specific tuor type, for which it is difficult to distinguish between benignity and alignancy. In the future, we want to deepen and expand our findings in several directions. We intend to use 5-fold cross-validation and ore rigid statistics on the 68- tuor data set, since it is very sall for training. A further challenge arises fro correlated tuors; which coe fro the sae patient. We intend to apply dedicated ethods for such instances within a bigger study in the future. Acknowledgeents. This work was supported by the DFG projects SPP 1335 Scalable Visual Analytics and SP 572/11-1 IMPRINT: Increental Mining for Perennial Object. References 1. Kuhl CK. The Current Status of Breast MR Iaging, Part I. Radiology. 2007;244(2): Chen W, Giger ML, Bick U, Newstead GM. Autoatic identification and classification of characteristic kinetic curves of breast lesions on DCE-MRI. Medical Physics. 2006;33(8): Glaßer S, Prei U, Tönnies K, Prei B. A visual analytics approach to diagnosis of breast DCE-MRI data. Coputers and Graphics. 2010;34(5): Prei U, Glaßer S, Prei B, Fischbach F, Ricke J. Coputer-aided diagnosis in breast DCE-MRI Quantification of the heterogeneity of breast lesions. European Journal of Radiology. 2012;81(7): D Rueckert, L Sonoda, C Hayes and et al. nrigid Registration Using Free- For Deforations: Application to Breast MR Iages. IEEE Trans Med Iaging. 1999;18(8): H Degani, V Gusis, D Weinstein and et al. Mapping pathophysiological features of breast tuors by MRI at high spatial resolution. Nat Med. 1997;3: Ester M, Kriegel HP, Sander J, Xu X. A Density-Based Algorith for Discovering Clusters in Large Spatial Databases with ise. In: Proc. of Knowledge Discovery and Data Mining (KDD-96); p Kailing K, Kriegel HP, Kröger P. Density-Connected Subspace Clustering for High- Diensional Data. In: Proc. of SIAM Conference on Data Mining; p Ankerst M, Breunig MM, Kriegel HP, Sander J. OPTICS: Ordering Points To Identify the Clustering Structure. In: Proc. of ACM SIGMOD Conference on Manageent of Data; p Holes G, Donkin A, Witten IH. WEKA: a achine learning workbench. In: Intelligent Inforation Systes,1994. Proceedings of the 1994 Second Australian and New Zealand Conference on; p Quinlan JR. C4.5: Progras for Machine Learning. Morgan Kaufann Publishers; 1993.