Lecce, April 13, 2011 – PhD discussion A Computer Assisted Detection system for juxta-pleural nodule identification in chest Computed Tomography images PhD tutors: Dr. Ivan De Mitri Dr.Giorgio De Nunzio PhD student: Andrea Massafra MAGIC-5 (Medical Applications on a GRID Infrastructure Connection) Detection (CAD) The Project isComputer funded byAssisted INFN - the Italian National Institute of Distributed Computing Infrastructure (GRID) Nuclear Physics - and coordinated with hospitals and Universities 6 Research Groups in Italy (Bari, Genova, Lecce, Napoli, Pisa, Torino, ~ 40 Researchers) Medical (Imaging) Applications: -Analysis of Digital Images MAGIC-5 Mammography- (breast cancer 2002-2006) Lung CT (nodules 2004 - present) Brain RMI (Alzheimer' s Disease 2006 – present) 2 MAGIC-5 research lines Mammography - automatic detection of masses and micro-calcifications in mammography Lung CT - detection of lung nodules in CT (Computed Tomography) images Brain RMI - early detection of Alzheimer’s disease in MR (Magnetic Resonance) and PET (Positron Emission Tomography) images PhD work Lung Segmentation algorithm 1 gray value threshold 2 region growing for lung reconstruction 3 wavefront algorithm for vascular tree extracion 4 lung fusion problem solving 5 masks extraction building of a CAD system for juxtapleural nodule detection,composed of a segmentation tool, a “nodule hunter” (a concavity-patching method such as the α−hull or morphological closing), feature calculation, and classification by an Artificial Neural Network implementation of conformal mapping algorithms for lung surface uniformization, as an alternative “nodule hunter” and a non-standard visualization tool fine tuning of the lung segmentation algorithm used in the MAGIC-5 Collaboration Materials Preliminary results from 57 CT scans from the MAGIC-5 DB: 78 juxta-pleural nodules, of which 28 directly in contact with the pleura (“pleural nodules”, easily detectable concavity in the segmentation mask!), the others connected by peduncles (“sub-pleural nodules”, small or almost absent concavity entrance). Internal Nodule: It originates in the internal part of the lung and it is fully embedded in the lung parenchyma. Sub-pleural Nodule: It originates in the internal part of the lung, but it can be found adjacent or connected to the pleural surface. Pleural nodule: It originates in the pleura and grow towards the lung parenchyma. 5 Structure of a CAD system Preprocessing - the aim of this stage is to reduce the amount of artifacts and noise in the image, enahancing its quality ROI detection the acronym of Region of Interest. In this step structures or regions in the image that contain possible pathological lesions are located - Structure/ROI Analysis - from each ROI, the system computes quantitative features that can characterize it (form, size and location, texture features, and so on) Classification - in the last step of CAD building, after each ROI is analyzed, it is individually classified as healthy or pathological. In CAD applications, classification is generally supervised and classifiers are often based on neural networks ROI detection Curvature Filtering mm^-1 Alpha hull : a convex hull generalization based on notion of shape of a point set Morphological closing: dilation followed by erosion of an image Structure/ROI Analysis Statistical texture analysis represents texture indirectly by non-deterministic properties that govern distribution and relationship between image gray levels First order statistical parameters measure the likelihood of observing a gray value at a randomly-chosen location in the image. They can be computed from the histogram of pixel intensities in the image. These depend only on individual pixel values and ignore the spatial interaction between image pixels. Second order statistical parameters are defined as the likelihood of observing a pair of gray values occurring at the endpoints of a segment of given length placed in the image at a random location and orientation. Geometrical Features: of the image under inspection. provides a good symbolic description Classification with Neural Network The k-th neuron consists of input signals having positive or negative weight associated with each input node the activation potential the activation function Structure of ANN the output signal Feed Forward Neural Networks patterns are presented to input units and the signal crosses the network from the input to the output layer Activation function Backpropagation algorithm The classification problem concerns the distinction between healthy and pathological subjects During network training, the output signal is compared with a target. The training error is The back-propagation learning rule is based on the gradient-descent optimization method. The algorithm stops when: fixed number of epochs is achieved or the classification error starts to increase on Validation Set (early stop method) or others conditions The ROC curve If the instance is really positive and it is classified as positive, it is called True Positive TP if the instance is positive, but it is classified as negative, it is called False Negative FN if the instance is negative and it is classified as positive, it is called False Positive FP if the instance is negative and it is classified as negative, it is called True Negative TN sensitivity specificity Juxta-Pleural nodule detection in CT images based on multiscale α-hull and morphological closing The procedure follows the classical scheme of a CAD system Juxta-pleural nodule candidate detection (ROI “hunting”) Lung segmentation closing Hierarchical tree Feature calculation Classification Alpha hull features: geometrical (span, depth…), texture (1° order) Conformal mapping 12 { Overview of segmentation algorithm A gray-value threshold for the segmentationm of the respiratory apparatus is performed by analyzing the image histogram 3D RG is applied to the CT volume. Voxels are included in the grown region if their Hounsfield Number is smaller than ϑ. The resulting binary mask containing the trachea, the bronchi, and the lungs. the external airways are extracted and removed by a wavefront simulation model with appropriate stop conditions. The resulting mask, containing only the lungs, is labeled M' partial volume effects reduction. The lungs from appear as a single object “fusion”). If that happens, the fusion is removed by threshold adjustment. ′′ Simple-threshold 3D RG is used, to grow the left and the right lung respectively, generating two masks with concavities (juxta-pleural nodules?) and vessels A method for concavity patching: α-hull In the Euclidean space a set B is said convex if The convex-hull of a point set I as the intersection of all the convex sets that contain I An α-ball b is an open ball with radius α the α−hull of a set S is defined as is the intersention of the complement of all the closed circles of radius 1/α that contain all the points of S. A methods for concavity patching: α-hull If α=∞ the α-hull is the convex hull of S If α=0 the α-hull is S itself A method for concavity patching: morphological closing the eroded set Y of a set X of points in space is the locus of centres x of Bx included in the set X where Bx is the structuring element and A method for concavity patching: morphological closing The dilation is defined as Where is the complement of the set The closing of an input image A by a structuring element B is defined as DETECTION OF NODULE CANDIDATES Concavities: found slice by slice by calculating the border difference between the original lung slice and the same treated by closing or α-hull : D(slice, α) = Aedge – Asm_edge D is a list of pixels belonging to the original border but not to the smoothed one, therefore it consists of a group of pixels identifying the concavities at the given value of α D Asm_edge Aedge – = NODULE mm Nodule radii histogram Asm_edge Aedge – D = α chosen: {8, 10, 12} mm FALSE POSITIVE 18 Juxta-pleural-nodule candidate detection: multiscale hierarchy By repeating, for a suitable set A of α values, difference operation and concavity search, a hierarchy of concavities (for each slice) is determined 1 0 .9 N001 N002 N003 N004 N005 N006 N007 N008 N009 N010 N011 N012 N022 N023 N024 N025 N026 N027 N028 N029 N030 N031 N032 N040 N041 N042 N043 N044 N045 N046 N047 N048 N054 N055 N056 N057 N058 N059 N060 N061 N070 N071 N072 N073 N074 N075 N076 N077 N013 N014 N015 N016 N017 N018 N019 N020 N021 N034 N033 N035 N036 N037 N050 N049 N051 N052 N053 N062 N064 N066 N063 N065 N067 N068 N069 N078 N080 N082 N079 N081 N083 N084 N085 0 .8 Each row corresponds to a scale level from the top (largest α) to bottom rows (smallest α) N038 N039 Concavities at scale α Tree implementation by an array 0 .7 0 .6 0 .5 0 .4 0 .3 N086 N087 N088 N089 N090 N091 N092 N093 N094 N096 N098 N095 N097 N099 N100 N101 N102 N103 N104 N105 N106 N107 N108 N109 N110 N112 N114 N111 N113 N115 N116 N117 N118 N119 N120 N121 N122 N123 N124 N125 N126 N128 N130 N127 N129 N131 N132 N133 Concavity n-ary tree (for a slice) N134 N135 N136 N137 N138 N139 N141 N143 N140 N142 N144 N146 N147 N148 N149 N150 N151 N152 N155 N153 N154 0 .2 0 .1 N156 N157 N145 If a segmentation boundary contains nested concavities, a set of α-hulls with successively larger values of α can be constructed to identify (by difference operation and concavity search) concavities at different scales as defined by α. This creates a natural hierarchy of concavities where the hierarchy is ordered by the α value, and allows tracking relationship between concavities. 19 Features Geometrical Features SPAN, DEPTH, BORDER LENGTH, AREA DEPTH/SPAN RADIUS CIRCULARITY SPAN Grey mean Textural Features: circularity Standard Deviation Where x is the mean Hounfield intensity distribution and Npix is the number of pixel DEPTH Features Skewness where xi is the gray value and p(x) is the frequency and б the standard deviation Kurtosis Shannon's Entropy where P(x) is the probability that X is in the state x and P lg2 P is defined as 0 if P = 0. Classification Classifier: supervised two-layer, 13 input, variable hidden neurons, 1 output feed forward ANN, trained with gradient descent learning rule with momentum k-fold cross validation (for us k = 3) From the ANN output on the whole dataset the Receiver Operating Characteristic (ROC) curves were drawn: AUC = 0.76, almost identical for mc and α-hull. Positive findings R O C c u r ve 1 0 .9 0 .8 1) Training: P1 U N1 U P2 U N2 - Test: P3 U N3 U Na3 2) Training: P1 U N1 U P3 U N3 - Test: P2 U N2 U Na2 3) Training: P2 U N2 U P3 U N3 - Test: P1 U N1 U Na1 s e n s it iv it y 0 .7 0 .6 0 .5 0 .4 0 .3 0 .2 0 .1 0 0 0 .2 0 .4 0 .6 1 - s p e c i fi c i t y 0 .8 1 22 Classification : number of neurons in hidden layer Early stop method 200 epochs Results: The prototype of a complete CAD system for juxta-pleural nodule detection based on the α−hull has been developed on a Dell T5500 Precision workstation (two quad-core Intel-Xeon [email protected] Ghz, 12 GB RAM) An ANN from the MatLab toolbox for neural network was used The whole system was implemented using MatLab environment The modular structure of the CAD system allowed the comparison of the α−hull efficiency with that obtained by Multiscale Morphological Closing the α−hull had better sensitivity (92.3% vs 84.6%) than morphological closing. The number of juxta-pleural nodules detected by the α-hull is 72 out of 78, while only 66 are collected by morphological closing (detection level) All the nodules lost by the α-hull are also lost by morphological closing Results: At classification level, the two methods proved roughly equivalent, with sensitivity and specificity around 72 − 75% and AUC around 0.75 − 0.77 Two different methods to stop ANN have been tested: • early stop • fixed number of epochs (200) From the computational point of view, the two methods are quite different: on an average CT scan (300 slices) • alpha hull nodule hunting 3 minutes for slice • morphological closing 10 minutes Paper 1)Giorgio De Nunzio, Eleonora Tommasi, Antonella Agrusti, Rosella Cataldo,Ivan De Mitri, Marco Favetta, Silvio Maglio, Andrea Massafra, Maurizio Quarta, Massimo Torsello, Ilaria Zecca, Roberto Bellotti, Sabina Tangaro, Piero Calvini, Niccolo Camarlinghi, Fabio Falaschi, Piergiorgio Cerello, and Piernicola Oliva, Automatic Lung Segmentation in CT Images with Accurate Handling of the Hilar Region, J0ournal of Digital Imaging ISSN0897-1889 (Print) 1618-727X (Online) OI10.1007/s10278-009-9229-11 (2009), 10- 20. 2)Giorgio De Nunzio, Eleonora Tommasi, Antonella Agrusti, Rosella Cataldo, Ivan De Mitri, Marco Favetta, Silvio Maglio, Andrea Massafra, Maurizio Quarta, Massimo Torsello, Ilaria Zecca, Roberto Bellotti, Sabina Tangaro, Piero Calvini, Niccolò Camarlinghi, Fabio Falaschi, Piergiorgio Cerello, and Piernicola Oliva, Automatic Lung Segmentation in CT Images with Accurate Handling of the Hilar Region, Journal of Digital Imaging ISSN0897-1889 (Print) 1618-727X (Online) DOI10.1007/s10278-009-9229-11 (2009), 10- 2. 3) G. De Nunzio, A. Massafra, R. Cataldo, I. De Mitri, M. Peccarisi, M.E. Fantacci, G. Gargano, E. Lopez Torres, Approaches to juxta-pleural nodule detection in CT images within the MAGIC-5 Collaboration, Journal of Digital Imaging, 24(2011), 11-27 Lectures Analisi delle Immagini Analisi statistica dei Dati Programmazione "Object Oriented" in C++ Tecniche fisiche per la biomedicina Modellistica Numerica e Analisi Dati Elenco delle Pubblicazioni Proceedings 1) DE NUNZIO, MASSAFRA A (Ieee Member).., E. TOMMASI, I.DE MITRI, R.CATALDO, M.FAVETTA, S.MAGLIO. (2008). An innovative lung segmentation algorithm in computed tomography images with accurate delimitation of the hilus pulmonis . NSS-MIC. Dresden, Germany. 2008. 2) G DE NUNZIO, MASSAFRA A.(Ieee Member)., L. MARTINA, R. CATALDO, S. MAGLIO, M. QUARTA, A. RETICO, L. BOLANOS. (2008). Lung Uniformization for Juxta-Pleural Nodule Detection. NSS-MIC. Dresden, Germany. 2008. 3) G. MERCURIO, S. MAGLIO, A. AGRUSTI, M. FAVETTA, MASSAFRA A(Ieee Member).., R. DEMITRI, A. CAVALLO. (2008). An information management system for distributed proteomic images: computational GRID technologies or a scale-free network of biobanks . NETTAB. Varenna, Como Lake. 2008. 4) G. MERCURIO, S. MAGLIO, A. AGRUSTI, MASSAFRA A(Ieee Member).., R. CATALDO, I. DE MITRI, M. FAVETTA A. MASSAFRA, G. MARSELLA, D. VERGARA, M. MAFFIA. (2008). Network P2P for exploring and visualization of proteomic data produced by two dimensional electrophoresis . The 21th IEEE International Symposium on Computer-Based Medical Systems. Jyvaskyla, Finland. June 17-19, 2008. (pp. 197-202). WASHINGTON, DC,: IEEE Computer Society (UNITED STATES). 5) R.CATALDO, M.QUARTA, A.AGRUSTI, G.DE NUNZIO, S.MAGLIO, M.E.FANTACCI, F.BAGAGLI, M.FAVETTA, MASSAFRA A (Ieee Member).., G.MERCURIO. (2008). Annotation of lung-screening images and 2D-E proteomic analysis for early diagnosis of lung cancer through federated biobanks . The Sixth International Conference on Bioinformatics of Genome Regulation and Structure. Novosibirsk, Russia. June 22-28, 2008 Proceedings 6) G.DE NUNZIO, S.MAGLIO, R. DEMITRI, A. AGRUSTI, R. CATALDO, I. DE MITRI, M. FAVETTA, G. MARSELLA, MASSAFRA A (Ieee Member)., M. QUARTA, AND G. MERCURIO (Ieee Member). (2008). Integrated Models for the Analysis of Two-Dimensional Electrophoresis Gel Images. Medical Imaging Conference. 7) Piergiorgio Cerello, Ernesto Lopez Torres, Elisa Fiorina, Chiara Oppedisano, Cristiana Peroni , Raul Arteche Diaz, Roberto Bellotti, Paolo Bosco, Niccolo Camarlinghi, Andrea Massafra Improving the Channeler Ant Model for lung CT analysis- Spie 2011 Workshop e Conferenze 1-2 Aprile 2008 Genova: pipeline lessons 3-10 june 2008 Alghero: School on software of nuclear physics 11-13 june 2008 Pisa: workshop on medical imaging 19-26 october 2008 Dresda(Germany): poster 19-20 jennuary 2009 Lecce: Meeting magic5 16-18 julay 2009 Pisa : Talk Comparison between methods for nodule inclusion 19-24 september2009 London (United Kingdom): MICCAI 2009 28 september 2009 Bari:SIF talk on lung segmentation algorithm 2 dicember 2009 Genova: Talk on Report of lung segmentation 12-13 may 2010 Lecce: workshop on medical imaging 18-19 september 2010: Pisa workshop on medical imaging 19-23 october 2010 Oporto( Portugal): TMSi conference Talk on comparison between closing and alpha hull