Paper 10574-21 Session 4: Machine Learning, 3:30 PM - 5:30 PM, Salon B Nearest Neighbor 3D Segmentation with Context Features Evelin Hristova, Heinrich Schulz, Tom Brosch, Mattias P. Heinrich, Hannes Nickisch Philips Research Hamburg, Digital Imaging February 11, 2018
Intro / Objectives Generality Automated 3D segmentation Supervised machine learning Speed Simplicity Binary context features Nearest neighbors classification Vantage point trees Scalability Accuracy Benchmark variations Speed up Data efficiency 2 Medical X-Ray Systems / Digital Imaging
Neaerest Neighbor Segmentation Pipeline Testing Training Training Features Test Features NNs Search Label Interpolation Extraction Extraction Labelling Post Processing Vantage Point Binary Context Trees Features 3 Medical X-Ray Systems / Digital Imaging
Binary Context Features Random selection of coordinates around central pixel Pairwise intensity comparisons τ p; x, y ∶= ቄ 1 if p(x) > p(y) 0 otherwise p(x i ) p(y i ) > ? p(x) is the intensity of p at a point x BRIEF Binary Robust Independent Repeat 𝑜 𝑒 times (e.g. 𝑜 𝑒 = 1280) , Elementary Features form a vector f : 𝑜 𝑒 LBP 2 𝑗−1 𝜐 𝑞; 𝑦 𝑗 , 𝑧 𝑗 𝑔 𝑜 𝑒 𝑞 : = Local Binary Pattern 𝑗=1 Capture contextual and structural information Computational efficiency (Hamming distance) Robustness to monotonic gray-level changes 4 Medical X-Ray Systems / Digital Imaging
Nearest Neighbor Search Vantage Point Trees (Construction) a 5 Medical X-Ray Systems / Digital Imaging
Nearest Neighbor Search Vantage Point Tree (Query) Efficient NN search for binary data Query k NN → Probability maps for each label 6 Medical X-Ray Systems / Digital Imaging
Experiments Feature Extaction NNs Classification Regularization Default (BRIEF + LBP) (Vantage Point) (Random Walker) Pipeline • Threshold • K-means • No regularization • Absolute • Kd tree Alternatives Difference • Median filter 70 abdominal CT images 42 pelvic MR images Liver, spleen, left kidney, right kidney Bladder, bones, prostate, rectum 512 x 512 x 394 (1.36 x 1.36 x 1.35 mm) 528 x 528 x 120 (1.04 x 1.04 x 2.5 mm) 5-fold cross validation (Train 56 / Test 14) 7-fold cross validation (Train 36 / Test 6) 7 Medical X-Ray Systems / Digital Imaging
Results, Dice Score Default No regularization Liver: 0.84, Spleen: 0.73, L. Kidney: 0.73, R. Kidney: 0.72 CT: 47% faster Bladder: 0.73, Bones: 0.63, Prostate: 0.61, Rectum: 0.64 MR: 27% faster 8 Medical X-Ray Systems / Digital Imaging
Results, Confusion Matrix CT MR Mainly correct predictions Few inter-organ confusions Often confusion with background (imbalanced training data) 9 Medical X-Ray Systems / Digital Imaging
Algorithm vs. Ground Truth (CT) Liver: 0.95 Spleen: 0.89 Left Kidney: 0.85 Right Kidney: 0.86 TP FN FP 10 Medical X-Ray Systems / Digital Imaging
Algorithm vs. Ground Truth (MR) Bladder: 0.91 Bones: 0.72 Prostate: 0.66 Rectum: 0.77 TP FN FP 11 Medical X-Ray Systems / Digital Imaging
In a nutshell.. Binary Context Features + Nearest Neighbors = Generic, simple, data-efficient, fast segmentation Generality Accuracy, room for improvement Dice: CT - 0.76, MR - 0.65 Speed Simplicity Scalability Accuracy Data efficiency 12 Medical X-Ray Systems / Digital Imaging
Training phase Training image Ground truth labels Body contour mask Sampling grid Storing features & labels BRIEF & LBP features extraction 14 Medical X-Ray Systems / Digital Imaging
Test phase BRIEF & LBP features extraction Storing test features VPF NNs query, retrieve labels Regularization Grid labels assignment Label interpolation 15 Medical X-Ray Systems / Digital Imaging
Recommend
More recommend
