Linee di ricerca: CAD polmonare, analisi di neuroimmagini. Sezioni INFN: Bari, Genova, Lecce, Napoli, Pisa, Torino. Contatti e collaborazioni: Ospedali [Genova (neuro, S. Martino), Lecce (lung, oncologico), Pisa (lung, AOUP e UNI), Roma (lung, Fatebenefratelli IT), Bracco Imaging SpA (Pisa, INFNMED, lung), IMAGO7 (Pisa, neuro)]. MAGIC-5 ( ) (Medical Applications on a Grid Infrastructure Connection) M.E. Fantacci
OUTLINE 2010 – LUNG: introduzione upgrades e test del VBNA CAD composizione CAM-RGVP-VBNA CAD news dal mondo radiologico: lung CAD, OsiriX sviluppi preliminari OsiriX evoluzione INFNMED – NEURO: studio di fattibilita’ di analisi di immagini MRI per la discriminazione del morbo di Alzheimer 2011 – Programma scientifico, milestones, manpower, richieste
Introduzione
VBNA CAD development: the goal Lung nodules are detected with two dedicated Computer-Aided Detection (CAD) procedures according to their location and shape CAD input CAD output CAD I for internal nodule detection CAD JP for juxtapleural nodule detection
CAD input CAD output VBNA CAD architecture: 1.Lung segmentation 2.ROI identification 3.ROI classification CAD I ROI z z-1 z+1 CAD JP ROI z z-1 z+1 VBNA CAD development: structure
3D morphological segmentation of the lung I.Isolation of parenchyma with a threshold on intensity and selection of the connected lung regions starting with a seed point inside each lung. Vessels and airways walls are not included in the segmented lung. II.A combination of dilation and erosion with spherical kernels r d and r e (with r e > r d ) followed by the logical OR operation between the mask obtained and the (a) mask is applied to include in the segmented lung all vessels and airways walls, while maintaining the original shape of the lung mask border I.II. initial volume segmented volume (a) VBNA CAD development: lung segmentation
The surface at the border between the lung parenchyma and the pleural wall has been segmented by an iso-surface triangulation technique using the marching-cube algorithm VBNA CAD development: lung segmentation
3D isometric array 11 nn ………… Convolution with 3D gaussians 11 12 13 ……… Eigenvalues of the Hessian matrix max = max [ 1 …… n ] 11 nn ………… Filter function n1 n2 n3 max is computed for each voxel Modeling the lung structures: Nodules → spherical shapes Blood vessels and airway walls → elongated shapes Fissures → planar shapes i = | i3 | 2 /| i1 | If i1 ≤ i2 ≤ i3 < 0 i = 0 otherwise VBNA CAD development: ROI hunter for CAD I
A peak detector algorithm finds the local maxima List of ROIs 59:293:226:5.0:0:peak1 54:308:213:5.0:0:peak2 175:251:215:5.0:0:peak3 363:249:142:5.0:0:peak4 50:252:243:5.0:0:peak5 323:175:173:5.0:0:peak6 371:150:128:5.0:0:peak7 … Example of a list of nodule candidates The list contains a large number of false positives 3D array of the Multi-scale Filter Function max FP nodule VBNA CAD development: ROI hunter for CAD I
P leural nodules have convex surface: the inward-pointing fixed-length surface normal vectors crossing the nodule surface intersect within the nodule tissue. 59:293:226:5.0:0:peak1 54:308:213:5.0:0:peak2 175:251:215:5.0:0:peak3 363:249:142:5.0:0:peak4 50:252:243:5.0:0:peak5 323:175:173:5.0:0:peak6 371:150:128:5.0:0:peak7 … A 3D matrix A(x,y,z) counts the number of surface normals that pass through the voxel (x,y,z) A(x,y,z) is smoothed with a gaussian to enhance the regions where many normals intersect. The local maxima in that matrix are collected in a list of nodule candidates VBNA CAD development: ROI hunter for CAD JP
Vector of “features” (i.e. sequence of grey-level intensity values of the 3D neighborhood of “v”) Neural classifyer Voxel “v” The basic idea: slice z slice z+1 slice z-1 3 eigenvalues of the gradient matrix (3x3 neighborhood) 3 eigenvalues of the Hessian matrix (3x3 neighborhood) + + (6 additional features) + VBNA CAD development: classification
Voxel-Based Neural Approach (VBNA) Voxels classified as nodule Voxels classified as normal tissue internal nodule juxtapleural nodule normal tissue From voxel classification to ROI classification: The trained neural network is applied to the voxels of each nodule candidate A nodule candidate is finally classified as “nodule” if the number of voxels tagged as “nodule” by the neural classifier is above a threshold VBNA CAD development: classification
Risultati del VBNA CAD sui diversi DB
Upgrades e test del VBNA CAD Riduzione “selettiva” di falsi positivi applicata al VBNA CAD per l’individuazione di noduli pleurici: – suddivisione dei FP in categorie – allenamento di un classificatore esperto per ogni categoria che sia in grado di distinguere i FP della sua categoria dai noduli
Upgrades e test del VBNA CAD
2010)
Upgrades e test del VBNA CAD
COMPOSIZIONE CAM-RGVP-VBNA CAD
CAM RGVP VBNA
COMPOSIZIONE CAM-RGVP-VBNA CAD Let pi; i = 1; : : : ; n denote the likelihood of each CAD nding. Every unique value of p in the set of n ndings corresponds to a point on the FROC curve of the system. For every unique p value we can compute the number of true positives TP when we consider all ndings with pi >= p as positive. We can also compute the number of false positives FP we obtain at this threshold (disregarding irrelevant ndings). Now we associate with each p a value where the factor +1 has been added in the denominator to avoid division by zero in the exceptional situation that all ndings are irrelevant, in which case both TP and FP equal zero. The values f(p) are approximately equal to the probability that a nding in the evaluation set with likelihood p or higher represents a true nodule. Such probabilities are natural measures to combine. To combine systems, we compute f(p) for every nding from every system. All findings are sorted so that we have fi; i = 1; : : : ; n and fi >= fj if i < j. Starting at fi with i = 1, it is checked for all ndings fj ; j = i + 1; : : : ; n if they correspond with fi. In this study we used the simple rule that ndings within 5 voxels of each other (and obviously located in the same scan) are corresponding.
COMPOSIZIONE CAM-RGVP-VBNA CAD A more elaborate criterion, such as the one used to compute the FROC curves in this study, could be used instead, but this is not possible as no segmentations or effective diameters of the input findings are available. If two findings fi and fj correspond, we set remove fj from the list of ndings and continue the procedure. It is easy to see that this is conceptually similar to averaging the probabilities for each nding across all systems, where undetected ndings correspond to a zero probability: we add up the ndings we are able to match across systems and if a system does not detect a particular nding, nothing will be added. Note that systems with low performance have f values that are close to zero for (nearly) all their ndings, and these systems are therefore automatically weighed less heavily in the combination.
COMPOSIZIONE CAM-RGVP-VBNA CAD
News dal mondo radiologico: lung CAD Presentati entusiasticamente a RSNA 2009 ed ECR 2010 i risultati di primi test CAD+RAD effettuati con lung CAD dalle prestazioni “alone” dichiarate ben peggiori dei nostri. Stazionaria la situazione sull’efficacia dello screening per la riduzione effettiva della mortalita’ per lung cancer (e di conseguenza le prospettive di utilizzo dei CAD in questo ambito). In rapidissima e positiva evoluzione le prospettive di uso dei CAD per i controlli oncologici (“screening” di metastasi polmonari, dose piena, possibilita’ di strato sottile). Lavoro sui CAD della collaborazione MAGIC-5: menzione d’onore al congresso SIRM 2010.
Pisa: intensivo uso di OsiriX presso il Dipartimento di Radiologia Diagnostica ed Interventistica, stretti contatti con gli sviluppatori, ampia rappresentanza in EuroPACS, contatto con Osman ECR2010 (PG & ME) News dal mondo radiologico: OsiriX
Sviluppi preliminari OsiriX Rad findings manager; Save/Retrieve rad findings; Cad findings import with probability selection; Cad findings manager; Hide/Show cad findings during rad annotation; PluginfeaturesPluginfeatures
Sviluppi preliminari OsiriX ESEMPIO
SVILUPPI INFNMED MAGIC-5 lung CAD Very good scientific results VBNA: 1° in the ANODE09 competition, poster award CARS09 and honorary mention SPIE 2010 MAGIC-5-lung “radiological work”: honorary mention SPIE 2010 Synergies between the different approaches (ANODE09 analysis) INFN-MED Requests (Bracco Imaging SpA: march 2009, im3d april 2009) OK from INFNMED committee: may 2009 Preliminary invitation letter (to 12 industries): february 2010 3 positive answers ( BioDigital Valley, BRACCO Imaging SpA, MEDICAD : march 2010 Planned: individual meetings with P. Cerello, R. Batistoni, A. Vacchi
MAGIC-5_neuro activities: analysis of MRI images to discriminate Alzheimer’s Diseases VBM (Voxel Based Morphometry) analysis of AD (Alzheimer’s Disease), MCI (Mild Cognitive Impairment), HC (Healthy Controls) subjects to identify significant local GM (Grey Matter) differences by means of a SPM (Statistical Parametric Map).
Attività in corso in neuro-imaging Automated localization of MTL ROIs (containing the Hippocampus) in MR T1 images Extraction of features and their analysis Identification of an image-based clinical index (physical observable) allowing for a reliable prediction on MCI patients Automatic segmentation of the hippocampus starting from few manually segmented templates Validation of the segmentation method by comparison with shapes manually segmented by expert readers [P. Calvini, A. Chincarini, G. Gemme, M.A. Penco, S. Squarcia, F. Nobili, G. Rodriguez, R. Bellotti, E. Catanzariti, P. Cerello, I. De Mitri, M.E. Fantacci, Automatic analysis of Medial Temporal Lobe atrophy from structural MRIs for the early assessment of Alzheimer disease, Med. Phys accepted for publication]
Neuroimages analysis
Neuroimages analysis: preliminary results
Neuroimages analysis: conclusions
Programma scientifico 2011 LUNG – Implementazione MAGIC-5-lung CAD e ottimizzazioni – Test CAD+RAD – Plugin OsiriX – Collaborazione con INFNMED NEURO – Estensione dell’analisi VBM con SPM alle immagini MRI strutturali del DB MAGIC-5 – Analisi whole brain – Sinergia con IMAGO7 per l’analisi di immagini con differenti caratteristiche
Manpower e richieste PISA 4 FTE “ortodossi”: – M. Barattini (spec. UNIPI) 100% – N. Camarlinghi (dott. UNIPI) 100% – S. Delle Canne (Fis. San. fbf-it) 100% – M.E. Fantacci (ric. UNIPI) 70% – A. Retico (ric. INFN) 30% Collaboratori (più o meno “sulle spese”): – I. Gori (Bracco Imaging SpA) – A. De Liperi, F. Falaschi (U.O. Radiodiagnostica 2 AOUP) – S. Angeli, D. Caramella, E. Neri, U. Tani, M. Zangani (dip. radiologia UNIPI) Richieste 2011: MI: 12 keuro (riunioni di collaborazione, ospedali ed esperti per test clinici, 2 congressi nazionali). ME: 10 keuro (tutorial e contatti OsiriX, 2 congressi internazionali). Consumo: 3 keuro (metabolismo). Inventariabile: 4 keuro (1 fisso e 1 portatile Apple per sviluppi Osirix e test clinici). Spazio: per noi e per i nostri computers. Servizi: calcolo (a livello ticket).