1. L'analisi delle immagini mediche per la diagnosi precoce delle neoplasie Roberto BELLOTTI Dipartimento Interateneo di Fisica M. Merlin Università degli Studi di Bari e Politecnico di Bari Istituto Nazionale di Fisica Nucleare - Sezione di Bari on behalf of the MAGIC-5 Collaboration

2. Main research activities Developing models and algorithms for a distributed analysis of biomedical images:  to support the radiologist's diagnosis with Computer-Aided Detection (CAD) algorithms  to improve computational speed, data accessibility and sharing of distributed image databases  to enable the co-working of medical experts and to allow large-scale statistical and epidemiological analysis Analysis of Medical Images:  Mammographic images for the early diagnosis of breast cancer (>2002)  Computed Tomography images for the early diagnosis of lung cancer (>2004)  MRI of the brain for the early diagnosis of the Alzheimer’s disease (>2006)

3. Interdisciplinary know-how Several techniques developed in High Energy Physics (HEP) and astrophysics experiments are implemented and optimized in medical image analysis to detect very low signal in a noisy background O. Adriani et al. (PAMELA Collaboration) Observation of an anomalous positron abundance in the cosmic radiation, NATURE, in print (2009)

4. Computer-aided detection of breast lesions

5. Image Acquisition & Manipulation (DICOM) Metadata & Diagnosis insertion CAD execution Data storage & retrieve through the GRID Operating Installations: Torino (Valdese), Lecce, Bari Hospitals Suzanne Mubarak Centre for Women's Health and Development, Alexandria (EGY) MASSES MICROCALCIFICATION CLUSTERS CAD station for Mammography Massive Lesion Microcalcifications [P. Cerello et al, Methods Inf Med 44, 244-248 (2005)]

6. Lesion Massives Recognition Segmentation: ROI (Region Of Interest) hunter Feature Extraction : co-occurrece matrix Classification: Neural Network

7. Breast CAD Area:0.783

8. Computer-aided detection of lung nodules in screening CT

9. Non-calcified small pulmonary nodules are considered as the primary signs of early-stage lung cancers Nodules with diameter ≥ 5mm have to be detected A CAD system could be useful as first or second reader It should be characterized by:  high sensitivity  low number of false-positive findings (FPs) per scan Lung CT Screening and Computer- Aided Detection (CAD) Thin-slice CT: Reconstructed slice thickness =1mm → ~300 slices/scan Low-dose helical multi-slice CT 0.6 mSv Low-dose helical single-slice CT 1.2 mSv Standard dose helical CT 5.0 mSv Rx torax 2 views 0.1 mSv

10. Lung nodules: examples The general CAD architecture: 1.Lung segmentation: defines the area where nodules are to be detected 2.Region of Interest (ROI) hunter: identifies a list of nodule candidates 3.ROI classification (False Positive findings reduction)

11. RG-ACM CAD: Regiong Growing & Active Contour Model 1. Lung segmentation: Region Growing + Active Contour Model 2. Nodule candidate identification: Region growing-based iterative algorithm 3. ROI classification: Rule-based filter + Neural Network [R. Bellotti et al, Med Phys 34,4901-1490 (2007)] Lung segmentation Nodule candidate identification False Positive reduction Internal force F i FiFi R i =F i+1 +F i+1 +F A Adhesive force F A q=F A /F i Rule-based filter Neural network

12. Lung CAD FROC curves Validation dataset: 24 CT (28 nodules with diameters ≥ 5mm ) 75% sensitivity @ 2÷6 FP/scan 0.75

13. Comparison with commercial systems R2 technology ImageChecker® CT Lung system (the first clinically validated CAD system for chest CT) 73% Sensitivity @ 3 FP/scan - 250 CT (140kV, 60mA, 1.25 mm slice thickness) [Roberts et al, CARS’05, pp 1137-1142] ImageChecker® v1.0, 56% Sensitivity @ 3.5 FP/scan - 30 MDCT (110 kV, 50-60 mA,1.25 mm slice thickness) [Brochu B, et al., J Radiol 2007;88:573-578] ImageChecker® (LM-1000) Sensitivity 73% @ 3.2 FP/scan - 150 MDCT (120 kV, 80 mAs, 2.5 mm slice thickness) [Yuan R, et al., Am J Roentgenol 2006;186:1280-1287] ImageChecker® (LM-1000) Sensitivity 60% @ 1.6 FP/scan - 70 MDCT (120 kVp, variable mAs, 2.5 mm slice thickness) [Lee IJ, et al., Korean J Radiol. 2005:89-93] Siemens Prototype of LungCAD CT 77.1% Sensitivity @ 2.7 FP/scan - 185 CT (120 kV, 90mAs, 1 mm slice thickness) [Wolf, CARS’05, pp 1143-1145] ICAD Az=0.72 - 18 MDCT (120 kV, 80 mAs, 0.75 mm slice thickness) [Marten K, et al., Eur Radiol 2004;14:1930-1938] Siemens vs R2 ImageChecker® Sensitivity 73% @ 6 FP/scan NEV Sensitivity 75% @ 8 FP/scan 25 MDCT (120 kV, 10-20-80 mAs, 1-2 mm slice thickness) [Das M, et al., Radiology 2006; 241:564-571]

14. Commercial and academic research systems The ANODE09 International competition for Nodule Detection in chest CT http://anode09.isi.uu.nl/ The results have been presented at SPIE Medical Imaging 2009

15. The early diagnosis of Alzheimer’s disease  Analysis of MR images of the brain  Evaluation of the atrophy of the hippocampus

16. What kind of information would be interesting for the neurophysiologist community? Early diagnosis by means of affordable and reliable tests “… a highly sensitive and specific diagnostic method for early detection of the disease is of the utmost importance for overall patient management and outcome.” years [Neurology 2002;59:198-205] years Overall cognitive performances for MCI subjects (arb. units) Mild Cognitive Impairment (MCI) predictors: only a fraction of the MCI population evolves in AD Evaluation of AD developers Initial conditions (all MCI) Performance decreases rapidly [34% of the initial population] Expected performance loss due to aging

17. Physical observables The goal is to find one (or more) observable, whose distribution:  maximizes the separation between Controls and AD population  is able to predict the evolution of a MCI patient Significant information is supposed to be encoded in the hippocampal ROI Hippocampal Boxes CSF/GM/WM Histogram equalization  Scalar product Median HB Controls Median HB AD [A. Chincarini et al, Computational Vision and Medical Image Processing, Tayolor & Francis, 121-126 (2007)] Database provided by: Ospedale S. Martino, Genova the Alzheimer’s Disease Neuroimaging Initiative (ADNI)

18. Analysis of MTL ROIs *

19. Conclusions Lung CAD  Good results  Integration of the different approaches  Database population Neuroimage analysis  Analysis of the hippocampus degree of atrophy for the early diagnosis of AD  Validation of the automated segmentation algorithm  Shape analysis  A physical observable computed on the hippocampal boxes can reliably predict the evolution of the MCI patients Contact: roberto.bellotti@ba.infn.it

20. Thanks for your attention! I am grateful to all the members of the MAGIC-5 Collaboration for their contribution