SemiAutomatic CentralChest Lymph Node Definition from 3 D

Semi-Automatic Central-Chest Lymph -Node Definition from 3 D MDCT Images Kongkuo Lu and William E. Higgins Penn State University Park and Hershey, PA, USA SPIE Medical Imaging 2010: Computer-Aided Diagnosis, San Diego, CA, Feb. 17, 2010

Motivation Ø Ø Ø Lung cancer is the leading cause of cancer death Early lung cancer detection could increase survival rate Diagnosis procedures: Ø 3 D MDCT chest image assessment Ø Follow-on diagnostic bronchoscopy 512 x ~700 voxels -1000 HU to 1000 HU (grayscale) Drawing by Terese Winslow, “Bronchoscopy, ” NCI Visuals Online, National Cancer Institute

Motivation Armato et al. , Radiology 2004 n n Subaortic and subcarinal lymph node (p 2 h 012 b) ROIs involve complex phenomena Manual slice tracing – not feasible Automatic methods – application dependent Semi-automatic methods – more practical Live Wire Lower para-tracheal lymph node (IRB 20349. 3. 3)

Prior Work 2 D Live Wire: Ø Ø Ø Mortensen and Barrett (Graphical Models and Medical Imaging 1998) Falcão et al. (Graphical Models and Medical Imaging 1998) Lu and Higgins (Int. J. CARS 2007) 3 D Live Wire: Ø Ø Ø Falcão and Udupa (Medical Image Analysis 2000) Hamarneh et al. (SPIE Medical Imaging 2005) König and Hesser (SPIE Medical Imaging 2005) Souza et al. (SPIE Medical Imaging 2006) Lu and Higgins (Int. J. CARS 2007) Poon et al. (SPIE Medical Imaging 2007 and CMIG 2008)

Prior Work Cost Function: Dynamic Graphic Search: 41 35 30 33 47 29 25 23 25 29 19 17 19 29 20 13 16 29 44 14 10 13 19 33 39 62 15 7 7 12 14 24 29 16 9 3 8 4 8 11 5 0 1 5 19 10 8 4 14 Mortenson GMMI 1998; K. Lu and W. E. Higgins, SPIE 2006 and IJCAR 2007 37 29 34 36 27 24 30 13 20 19 22 33 11 13 14 20 28 40 15 18 22 24 29

Prior Work Segment ROI using 3 D 2 D Live Wire K. Lu and W. E. Higgins, SPIE 2006 and IJCAR 2007

Central-Chest Lymph Nodes in 3 D MDCT Image Trans Case IRB 20349. 3. 3 M 4 -1 M 4 -2 Coro M 4 -1 M 4 -3 Complete segmentation of Typical lower paratracheal (M 4) lymph nodes central-chest lymph nodes

Single-Section Live Wire

Single-Section Live Wire Ø Adjust working area adaptively Ø Refine seed set iteratively Ø Terminate iterations when Ø Terminate 3 D process when Ø Section limits reached Ø Boundary costs vary greatly Ø Boundary is too small Ø Segmented Ø Intensity region is too small distribution varies greatly K. Lu, Ph. D Dissertation 2010

Single-Section Live Wire

Single-Section Live Wire

Single-Click Live Wire

Single-Click Live Wire

Single-Section and Single-Click LW Single-Section Single-Click Lu, Ph. D Dissertation 2010 and K. Lu and W. E. Higgins Computer in Biology and Medicine (in submission)

Performance Evaluation MDCT Data: Computer - Dell Precision 650 workstation: • Dual Intel Xeon 3. 2 GHz, 3 GB RAM, Windows XP

Results Method Number of Nodes Success Rate Accuracy Inter-Trial Reproducibility Processing Time Segmented Nodes: q Short-Axis Length: q Long-Axis Length: q Volume: q Number of Voxels: Single-Section Single-Click 50 90 % (45/50) 80% (40/50) 81± 7% 79± 8% 88± 7% 16± 4 s 5. 8± 1. 5 mm 11± 4. 0 mm 256± 210 mm 3 1089± 861 86± 9% 20± 5 s

Comparison of Two Observers Accuracy Reproducibility Processing Time 20349. 3. 15 O 1 O 2 83% 80% 87% 88% 21405. 67 O 1 O 2 78% 79% 85% 16 s 19 s Operator Independent 20 s 22 s

Example Lymph Node Segmentations 21405. 64 * Segmentable nodes: both observers and both methods 20349. 3. 3

Summary Single-Section and Single-Click Live Wire Ø Reduce human interaction Ø Efficient and reliable Ø Operator independent Ø Can handle typical lymph nodes and other ROIs • 90% Single-Section • 80% Single-Click

Acknowledgments NIH NCI grant #R 01 -CA 074325 • Pinyo Taeprasartsit - experiment participant • The Multidimensional Imaging Processing Lab at Penn State

Single-Section and Single-Click LW Evaluation Lymph node #: Succ. Seg. : Accuracy: Inter-trial Reprod. : 50 M 1 -45 (90%) >80% >85% M 2 -40 (80%) M 1, M 2: Single-section and single-click LW; ε 1, ε 2: Trials; a(Mi, εj): accuracy; r(Mi, εj, εk): reproducibility

Single-Section and Single-Click LW Evaluation M 1, M 2: Single-section and single-click LW; ε 1, ε 2: Trials; a(Mi, εj): accuracy; r(Mi, εj, εk): reproducibility