Known causes: Drugs Collagen Vascular Exposures Genetict IIPs: Idiopathic interstitial pneumonias Granulomatous lung diseases: Sarcoidosis Fungal Mycobacterial Unique entities: Pulmonary Alveolar Proteinosis Eosinophilic Granulomatosis Eosinophilic Pneumonia Lymphangioleiomyomatosis Capillaritis IPF: idiopathic pulmonary fibrosis 55 % 25 % NSIP: nonspecific interstitial pneumonia RB-ILD: respiratory bronchiolitis interstitial lung disease 10-15 % <2% AIP: acute interstitial pneumonia DIP: desquamative interstitial pneumonia 5% COP: cryptogenic organising pneumonia LIP: lymphocytic interstitial pneumonia <1%
CT Scan Lung Segmentation Lung Field Segmentation Bronchovascular Tree Segmentation ILD Pa ern Classification Feature Extraction Feature Selection Classification ILD antification ILD map Findings, Extend, Localization Diagnosis Rule Based Scheme Machine Learning Clinical Parameters
R i,j i j
C i,j i j 2 8
D n D D
n k k n
X = [x 1, x 2,, x n ] C j C = [c 1, c 2,, c k ] p(c j x 1, x 2,, x n ) p(x 1, x 2,, x n C j )p(c j ) p(c j x 1, x 2,, x n ) X C j n p(x 1, x 2,, x n C j ) p(x k C j ) n p(c j x 1, x 2,, x n ) p(c j ) p(x k C j ) X C j k=1 k=1 x n b y( x) = f(w T x) = f( n i=1 W ix i + b) f : R R W W y( x) x
X 1 H 1 X 2 Y 1 X 3 H n Y n X n ϕ
φ S S {h(x, Θ k ), k = 1,... } {Θ k }
0 u, v < 4 [0, 1] I(x, y) F (u, v) F (u, v) = α(u)α(v) N 1 N 1 x=0 y=0 I(x, y) [ π(2x + 1)u π(2y + 1)v ] [ ] 2N 2N
u, v = 0, 1,..., N 1 α 1/N x = 0 α(x) = 1/N x 0 I(x, y) = N 1 N 1 u=0 v=0 α(u)α(v)f (u, v) [ π(2u + 1)x π(2v + 1)y ] [ ] 2N 2N F (u, v) I(x, y) α(u)α(v) [ π(2u + 1)x π(2v + 1)y ] [ ] 2N 2N 5 5 0 u, v < 4 f uv (x, y) u, v [1, N 1] N N 0 u, v < N 1 {I u,v (x, y)} I(x, y) X F X (x) = (X x) X x x
Convolute original image with filter bank Feature vector Original image Filtered images q-quantiles X1 X2 Filter bank Intensity histogram Xn F 1 (y), y [0, 1] x F (x) = y q q 1 k 0 < k < q D = (N N) q + b q b [0, 1] 5 5
n n n
n
Feature Vector Fixed Scale Classifier RF classifier model Classifier s votes V1 V2 Classifier s votes in Patches V 2 V 1 V 2 V 1 V n Vn 100 90 80 100 90 80 70 60 50 40 30 20 10 0 100 90 80 70 60 50 40 30 20 10 0 70 60 50 40 30 20 10 0
Σχήμα 3.5: Παραδείγματα ταξινόμησης πνευμονικού παρεγχύματος. Από αριστερά προς τα δεξιά: μέρη αξονικής τομογραφίας, ενδείξεις προτύπου αναφοράς, πρόβλεψη του υπολογιστή. Από πάνω προς τα κάτω: φυσιολογικό (κόκκινο),εσμυρισμένο (κίτρινο), κυψελωτό (μπλε), ενοποιημένο (πράσινο), δικτυωτό (κυανό), δικτυωτό/εσμυρισμένο (μοβ). 23
increased attenuation tuberculosis early fibrosis peripheral micronodules cysts macronodules bronchiectasis pcp emphysema fibrosis reticulation healthy consolidation ground glass micronodules 1mm
1 2mm 10 15mm 1 2s 512 512 0.4 1mm ±
db_root ILD_DB_lungMasks Case_id lung_mask lung-mask-0001.dcm CT-0001.dcm ILD_DB_volumeROIs Case_id roi_mask roi-mask-0001.dcm CT-0001.dcm ILD_DB_txtROIs Case_id CT-txtROIs.txt CT-0001.dcm case_id: '109' label: 'ground_glass' localisation: basal slice_number: 11 xvalues: [1x134 double] yvalues: [1x134 double] spacing_x: 0.6836 27
Study: 1.2.124.113532.129.195.3.60.20030403.100010 Series: 1.2.840.113704.1.111.154.1049453263.6 SpacingX: 0.68359375 SpacingY: 0.68359375 SpacingZ: 10.0 label: ground_glass localisation: basal slice_number: 9 nb_points_on_contour: 180 252.86186026506311 158.70507094197274 254.27775183383142 156.58123358882034 255.3396705104076 154.45739623566791 255.69364340259966 153.04150466689964 256.7555620791759 151.27164020593932 257.1095349713679 150.2097215293631 257.46350786356004 149.1478028527869 258.1714536479442 147.37793839182655 258.8793994323283 145.25410103867412 259.23337232452036 144.19218236209792 259.58734521671244 142.42231790113755 259.9413181089045 141.0064263323693 259.9413181089045 139.23656187140892 259.9413181089045 138.17464319483273 spacing_y: 0.6836 ct_scan: [1x83 char] lung_mask: [1x99 char] roi_mask: [1x98 char]
21 21 90
initial database new database re-structure patch extraction patches dataset database browser tool Talisman db_browser GUI re-annotate MAT-file F avg = 1 M M F c c M c=1 F c F c = 2 precision c recall c precision c recall c recall c = samples correctly classified as c samples of class c precisson c = samples correctly classified as c samples classified as c
Features PIXLV GLDM HIST QUANT GLRLM GLCM LBP 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 F-score (%) F avg ( )
F avg ( ) 3 3 k = 10 k = 10
F avg ( ) F avg ( ) F avg ( )
F-score 0.9 0.89 0.88 0.87 0.86 0.85 0.84 0.83 Forward SFS Backward SFS no SFS 0.82 0 2 4 6 8 10 12 14 number of feature selection steps
5 5
Properties: db_root: 'C:\{$db_root}' directories: [120x1 struct] cases: [120x1 struct] rois: [2084x1 struct] Methods: setdbroot: sets the db_root property getalldirectories: scan db_root for all subdirectories and store them at directories property loadrois: loads all ROIs at rois property loadcases: loads all CASEs at cases property filterrois: filters ROIs using pattern labels filtercases: filters CASEs using pattern labels createpath: return the actual path of ct scans by combining the db_root with [directories] id: {Current case id} age: {Age of patient} gender: {Gender of patient} rois: [1x3 struct] directories: [1x1 struct]