Presentation on theme: "Tumor Discrimination Using Textures Presented by: Maysam Heydari."— Presentation transcript:
Tumor Discrimination Using Textures Presented by: Maysam Heydari
Introduction Main goal: Discrimination between different tumor grades/types using textural properties Tumor pathologies: –Grade 2: astro (7), oligo (22) –Grade 3: aa (2), ao (1), amoa (1) –Grade 4: gbm (17)
Introduction Patient data: –50 unique patient-study pairs: 25 expert segmented patients 25 Maysam segmented patients –For each patient, the study nearest to the biopsy date (in the range ±52 weeks) was picked. –The nearest biopsy was chosen to determine the pathology.
Weeks between study and biopsy Expert segmented weeks # of patients Maysam segmented (low grade tumors)
Texture Features Features extracted on the segmented tumors: ENH (T1, T1C) and EDE (T2) on every slice. Each pixel in the tumor receives a texture intensity: –Gray Level Co-occurrence Matrices (GLCM) –MR8 –BGLAM left-to-right symmetry similarity values
Texture Features GLCM stat measures: –Energy: “orderliness” of pixels –Contrast:
Texture Features MR8 filter bank: For each filter, max response over 6 orientations Filters: –3 scales of edge filters –3 scales of bar filters –A Gaussian –Laplacian of Gaussian
Texture Features BGLAM: –Texture similarity of the segmented tumor to the symmetric side of the brain.
T1T1CT2 raw energy contrast BGLAM simvals Patient: 145 Study: 2
Method Each patient’s tumor is represented by a histogram for each modality and texture feature. The histograms are used as vector inputs to kmeans (k = 2) clustering.
Test Results lowgrade/highgrade: mismatch rates T1 T1C T2 Raw st MR nd MR rd MR th MR th MR th MR th MR th MR Energy Contrast BGLAM
Test Results gbm/rest: mismatch rates T1 T1C T Raw 1st MR8 2nd MR8 3rd MR8 4th MR8 5th MR8 6th MR8 7th MR8 8th MR8 Energy Contrast BGLAM
What’s Next? Combine the histograms from several texture features … –Stack them as vectors? Curse of dimensionality … with only 50 data inputs. –Instead of histograms, use stats: mean, var, min/max? Supervised learning –SVM?