Department of Electrical Engineering and Computer Science

1
Automatic Segmentation of Brain Structures for
Radiation Therapy Planning
Pierre-François DHaese, Valerie Duay, Rui
Li Aloys du Bois dAische, Thomas E. Merchant,
Anthony J. Cmelak, Edwin F. Donnelly, Kenneth J.
Niermann, Benoit Macq, Benoit M. Dawant

• Department of Electrical Engineering and Computer
  Science
• Vanderbilt University, Nashville, Tennessee

SPIE 2003 Medical Imaging
2
Contents

• Introduction
• Method
• Atlas Generation
• Atlas based segmentation
• Contour correction
• Validation Method
• Results and Discussion
• Conclusion Future work

SPIE 2003 Medical Imaging
3
Introduction

• Objective
• To provide accurate and robust automatic
  segmentation of selected structures of the brain
  on MR images volumes for computer aided
  radiotherapy

SPIE 2003 Medical Imaging
4
Atlas Based Segmentation

• Concept

T
ATLAS

• Rigid Registration
• Non Rigid Registration
• Model projection

SPIE 2003 Medical Imaging
5
Atlas Based Segmentation

• Concept

ATLAS

• Model projection

SPIE 2003 Medical Imaging
6
Data Sets Two Data Sets Different challenges

• Adults Data Set (20 patients)
• Adults
• Visible pathologies
• (unresectable glioblastoma multiforme to
  prepontine memingioma)
• Voxel resolution
• 1 mm³ isotropic
• Children Data Set (45 subjects)
• Children/Young Adults (aged between 1-21)
• Visible Pathologies
• (Infratentorial ependymoma)
• Poor voxel resolution
• 0.78² by 3 mm
• Large lesions or abnormalities

SPIE 2003 Medical Imaging
7
Atlas Based Segmentation

• Concept
• Model projection using deformation computed
  using image registration
• Rigid Body Registration
• Based on Mutual Information
• (Maes et al , Viola et al. 1996 )
• Non Rigid Registration
• ( Rohde et al. 2003)
• Deformation field ? linear combination of radial
  basis functions
• Use of Mutual Information
• Irregular grid
• Multi Resolution

SPIE 2003 Medical Imaging
8
Atlas Generation

• Atlas Selection and Structures delineation
• Most representative subject
• Manual Delineation of structures by a radiation
  oncologist
• Smoothing and mesh generation
• Discontinuities between slices
• Drawing irregularities
• 3-D Spline based smoothing of the contours

SPIE 2003 Medical Imaging
9
Correction of contours Cerebellum mis
registration

• Difficulties for Cerebellum
• Poor edge definition
• Correction using Level Sets method. (Sethian et
  al.)
• Contour shrunk into the cerebellum
• Speed function prevents leakage into adjacent
  structures (brainstem, cortex)

SPIE 2003 Medical Imaging
10
Validation Manual Contour Delineation

• Validation on randomly selected slices
• 3 Raters
• 1 Radiation Oncologist (A. Cmelak)
• 1 Radiologist (Ed. Donnelly)
• 1 Junior physician (K. Niermann)
• 5-8 randomly selected slices per structure
• 9 Structures (brainstem, cerebellum, chiasm,
  eyes, lenses, optical nerves)
• 11 patients (Adults data set)
• ? 2046 manual contours

SPIE 2003 Medical Imaging
11
Validation

• Qualitative and visual validation
• Quantitative validations

• Contour Based Validation
• Definition of an Envelope
• points IN the envelope
• Points OUT
• Mean Distance Error
• Max Distance Error

• Mask Based Validation
• Similarity Measure
• N(x) nb of points for mask x

0 (no matching) S 1 (perfect matching)
SPIE 2003 Medical Imaging
12
Results and Discussion (1/8) Adults
SPIE 2003 Medical Imaging
13
Results and Discussion (2/8) Adults
SPIE 2003 Medical Imaging
14
Results and Discussion (3/8)
SPIE 2003 Medical Imaging
15
Results and Discussion (4/8)
SPIE 2003 Medical Imaging
16
Results and Discussion (5/8) Contours similarity
( pts in envelope) per patient
SPIE 2003 Medical Imaging
17
Results and Discussion (6/8) Mask similarity
per patient
SPIE 2003 Medical Imaging
18
Results and Discussion (7/8) Optical Nerves
Challenging structures

• Optical Nerves
• Clearly visible close to the eyes (black)
• White when connecting to the chiasm
• Difficult to delineate manually
• Need of a priori structure information

SPIE 2003 Medical Imaging
19
Results and Discussion (8/8) Children/Young
adults Some challenging cases
SPIE 2003 Medical Imaging
20
Conclusion - Future Work

• Large Structures
• Computed Aided Radiotherapy is a possibility
• Small structures ? Still challenging
• Ultimate goal
• Comparison of radiotherapy plans made using
  automatic and manual contours.

SPIE 2003 Medical Imaging