Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11....

35
Spike Detection (1D)

Transcript of Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11....

Page 1: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Spike Detection (1D)

Page 2: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 3: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 4: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Spike detection - Simple Threshold• Null: No preprocessing, just apply a positive threshold to raw data• Negation: Signal is inverted before threshold. This is same as applying a negative threshold• Absolute Value: Same as applying both positive and negative thresholds simultaneously

Page 5: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 6: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 7: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Spike detection – Adaptive Threshold

• Ex. set an automatic amplitude threshold

40.6745x

T medianì ü

= × í ýî þ

Page 8: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Spike detection – Energy Based

• NEO: Non-linear Energy Operator = − + − ,ℎ 1 ≤ ≤ 4• SNEO: Smoothed NEO

• Convolution of NEO with a smoothing window function

Page 9: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 10: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 11: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Spike detection – Matched Filter (MF)

• MF: raw signal convolved with a template• Template is an average of action potentials• Could be NEO + MF, Absolute Value + MF, etc

Page 12: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 13: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Spike detection – Wavelet

Page 14: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

• Two different wavelet-based approaches presented in this work: (a) Wavelet multiresolution analysis is composed by wavelet decomposition and reconstruction through low-pass (LP) and high-pass (HP) filter banks; (b) How to map the wavelet scalogram

Page 15: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

• SWT = Stationary Wavelet Transform

Page 16: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

DL Object Detection (2D)

Page 17: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

CNN Applications• Classification• Detection• Segmentation• …

Classification

Page 18: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Regional Convolutional Neural Networks (R-CNN)

18

Girshick et al., Rich feature hierarchies for accurate object detectionand semantic segmentation, CVPR, 2014

① R-CNN② Fast R-CNN③ Faster R-CNN④ Mask R-CNN

Object Detection and Classification

Page 19: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

R-CNN의 기본 프로세스

• Selective search 알고리즘을 사용하여 2000개의 region proposal을 추출• Selective Search: Selective Search for Object Recognition, J. R. R.

Uijings et al, IJCV13.• 모든 regional proposal 검출 영상을 같은 image size로 변환• CNN을 통하여 특징 추출 (forward propagation)• SVM (Support Vector Machine)을 통하여 추출한 특징을 분류하여 object classification• Per regional proposal image, CNN runs. Inference speed is slow. Also multiple stages of region proposals, CNN feature extraction, and classification via SVM.

Page 20: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

You Only Look Once (YOLO)

20

- Real-time Object Detection- Also Classification

① YOLO v1.0② YOLO v2.0③ YOLO v3.0

Page 21: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Deep Learning YOLO

21

- YOLO = You Only Look Once- Real-time Object Detection- Joseph Redmon, Ph.D candidate,

University of Washington, CSE

Page 22: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

YOLO 특징

• R-CNN은 detection을 위하여, regional proposal을 추출하고, classifier 사용• YOLO는 object bounding boxes와 class probabilities를 regression 방법으로 예측• 전체 구조가 single neural network frame• 장점

• Extremely fast: detection을 regression 방식으로 해결• YOLO sees whole image, not regions• YOLO can be general, working on all kinds of images

• 단점• Fast at detection, little weak at finding position of small objects

Page 23: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 24: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 25: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 26: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

How YOLO Works1. Divide an input image into SxS grid cells2. grid cells들은 B개의 bounding box와, 각 bounding box에 대한 confidence score를 계산

Confidence Score = 3. 각 grid cell은 C개의 conditional class probability를 가짐.

Conditional Class Prob. = Pr(Classi | Object) 4. 각 bounding box는 x, y, w, h, confidence로 구성

(x,y): bounding box의 중심점(w,h): width & height

5. 각 grid cell은 (B*5+C) vector를 가짐. B*5: B개의 bounding box + x, y, w, h, confidence 즉 5. C는 class의 수

6. Each whole image has (S*S*(B*5+C)) tensor. 정보 기록7. At test, compute class-specific confidence score

Class Specific Confidence Score =

Page 27: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

What is IOU?

Page 28: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 29: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive
Page 30: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Medical Computer Aided Diagnosis (CAD) System

Page 31: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Deep Learning Model: Detection & Classification:

31

YOLO-based Medical CAD

Page 32: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

YOLO-based Medical CAD System• Detection Classification

Page 33: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Deep Learning Segmentation

33

Full Resolution SegNet

Page 34: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Deep Learning Models: Detection, Segmentation, & Classification

34T.-S. Kim et al., A fully integrated computer-aided diagnosis system for digital X-ray mammograms via deep learning detection, segmentation, and classification, Int. J. of Medical Informatics, 2018

Page 35: Spike Detection (1D)web.khu.ac.kr/~tskim/NE Lect 19-1 Signal Detection - Deep... · 2018. 11. 12. · Spike detection -Simple Threshold •Null: No preprocessing, just apply a positive

Deep Learning Models: Detection, Segmentation, & Classification

35