video summarization of key events stage i - the critical view michael a. grasso, md, phd university...

37
Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science MichaelGrasso.com

Upload: morgan-davidson

Post on 26-Dec-2015

221 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Video Summarization of

Key EventsStage I - The Critical View

Michael A. Grasso, MD, PhDUniversity of Maryland School of

MedicineUMBC Computer Science

MichaelGrasso.com

Page 2: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Abstract

Laparoscopic surgery is a minimally invasive technique with unique training requirements. Video-assisted evaluation is one method that surgical residents can use to demonstrate competence. Automated video summarization can increase the efficiency of evaluations by directing the senior surgeon to key portions of a surgical procedure. We are using image classification techniques to segment videos of laparoscopic cholecystectomies to assist with surgical training and evaluation.

Page 3: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Overview

Background Laparoscopic Surgery Image Classification

Methods Discussion

Page 4: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Laparoscopic Surgery

Minimally Invasive Surgery. First performed in 1987. Used in many surgical procedures.

Gall bladder removal (cholecystectomy).

Esophageal surgery (fundoplication). Colon surgery (colectomy). Others.

Page 5: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Laparoscopic Approach Narrow tubes (trocars)

are inserted into the abdomen through small incisions.

www.fda.gov

Page 6: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Laparoscopic Procedure Camera is passed

through trocar. Procedure is often

videotaped. Carbon dioxide is

infused through trocar.

Instruments are passed through the trocars to cut, manipulate, and sew.

Page 7: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Laparoscopic Aftercare Compared with an

open procedure. Smaller scars. Reduced pain. Quicker recovery.

http://www.nlm.nih.gov/medlineplus/ency/presentations/100166_1.htm

Page 8: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Technical Challenges

Access limited to small incisions. Long instruments with only the tips

visible. Two-dimensional video. Limited tactile feedback.

British Journal of Surgery. 2004 Dec;91(12):1549-1558

Page 9: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Laparoscopic Training

Traditional apprenticeship model. Acquire skills during actual procedures. Not sufficient for laparoscopic skills.

Other methods. Box trainer with animal or synthetic

models. Virtual reality simulator. Video-based assessment.

Page 10: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Assessment of Skills

Trainee must demonstrate competency.

Evaluation by a senior surgeon. Direct observation of the trainee. Video-based assessment.

Question: Can we organize video in order to assist in video-based assessment?

American Journal of Surgery. 1991 Mar;161(3):399-403

Page 11: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Objective

Identity key portions of surgical procedure to aid in video-based assessment.

Stage I is to identify the "critical view".

Video

Segments

Frames

Page 12: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Overview

Background Laparoscopic Surgery Image Classification

Methods Discussion

Summary: Organize surgical video to make it easier for expert to review.

Page 13: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

The Critical View

Helps ensure that the anatomy has been properly identified.

Occurs after dissecting anatomy. Occurs before clipping the cystic

artery and cystic duct.

Page 14: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

The Critical View

Cystic artery

Liver

Cystic duct

Fundus

Netter's Atlas of Human Anatomy

Page 15: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

The Critical View

Page 16: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Image Classification - Human

Features a person might use. Spectral features.

Tonal variations. Textural features.

Spatial distribution of tonal variations. Contextual features.

Features from surrounding areas.

Page 17: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Image Classification - Computed

Features extracted from image. Spectral features.

Distribution, size, width. Textural features.

Homogeneity, contrast, correlation. Similarity/distance metrics.

Jaccard coefficient, Jeffrey divergence.Journal of WSCG. 2003; 11(1):269-273

IEEE Transaction on Systems, Man, and Cybernetics. 1973 Nov; 3(6):610-621

Page 18: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Color Histogram Red, green, blue, or gray.

Count number of pixels for each tone. One 28 set for an 8-bit image for each color. Does not vary with translation and rotation. Ignores shape and texture.

4x4 image. 4 gray tones. H = {5, 4, 5, 2}

0 0 1 1

0 0 1 1

0 2 2 2

2 2 3 3

Page 19: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Binary Histogram

Quantize values for each tone to 0 or 1.

Background color given less weight. Subtle changes given more weight.

HB = {1, 0, 1, 1}

0 0 0 0

0 0 0 0

0 2 2 2

2 2 3 3

Page 20: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

3D Histogram

Distribution within a 3D color-space. 3D color space (red, green, blue). Used in object recognition & image

retrieval. n3 entries, where n = number of tones.

Example. Quantized to 3 tones

for each color.

Page 21: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Spatial-Dependency Matrix Co-occurrence

matrix. Co-occurring values

(0o, 45o, 90o, 135o). Four 28 x 28

matrices for 8-bit image.

Co-occurring Bits

0 1 2 3Refe

ren

ce B

its

0 4 2 1 01 2 4 0 02 1 0 6 13 0 0 1 2

0 0 1 1

0 0 1 1

0 2 2 2

2 2 3 3

135o 90o 45o

0o Ref 0o

45o 90o 135o

M0 =

Page 22: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Additional Spectral Features

Location of the distribution. Mean = Σ (bin*freq) / Σ (freq). Mode = bin of the max freq.

Size of the distribution. Standard deviation.

Width of the distribution. Max(bin) - Min(bin).

Page 23: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Additional Textural Features

Homogeneity. Number of tone transitions.

Contrast. Amount of local variation.

Correlation. Measure of linear dependencies.

Page 24: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Similarity/Distance Metrics

Jaccard Coefficient. Similarity of two sample sets.

|A B| / |A B| Two binary sets.

M11 / (M01 + M10 + M11)

Jeffrey Divergence. Distance between two vector spaces.

Σ (xi log(xi/avgi) + yi log(yi/avgi))n

i=1

Page 25: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Other Distance Metrics

City Block or Manhattan Distance. Euclidean Distance. Chi-Square. Canberra Distance.

Proceedings ACM SAC. 2008;:1225-1230

Page 26: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Related Efforts - Hysteroscopy Use Jeffrey divergence on color

histogram to identify segments. Relevant segments based on image

redundancy. No understanding

of the content of each segment.

Proceedings 27th IEEE-EMBS. 2005;:5680-5683

Mayo Clinic

Page 27: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Related Efforts - Echocardiogram

Use cosine similarity and edge change ratio to identify video segments.

State-based modeling. Identify states in each

video segment. Diastole (resting). Systole (contracting).

IEEE Transaction on Information Technology in Biomedicine. 2008 May;12(3):366-376

Medline Plus

Page 28: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Overview

Background Laparoscopic Surgery Image Classification

Methods Discussion

Summary: Spectral and textural features compared with similarity metrics.

Page 29: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Methods

Our objective. Identity key portions of surgical

procedure to aid in video-based assessment.

Stage I is to identify the "critical view".

Video

Segments

Frames

Page 30: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Tools

FFmpeg http://ffmpeg.mplayerhq.hu/ Extract JPEG images.

ImageJ http://rsbweb.nih.gov/ij/ Macros and Java plugins.

Page 31: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Work Plan

Identify videos for analysis. Convert videos to JPG. Evaluate ability to identify critical view.

Color histogram. Binary histogram. 3D histogram. Spatial-dependency matrix. Jaccard coefficient, Jeffrey divergence.

Page 32: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Algorithm

Feature ExtractionImageJ Color Histograms

Binary Histograms3D Histograms

Spatial-Dependency Matrices

Similarity Metric

Critical View?

Critical View

Random Image

Image ExtractionFFmpeg

Page 33: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Overview Background

Laparoscopic Surgery Image Classification

Methods Discussion

Summary: Attempt to identify the critical view by comparing image features with similarity metrics.

Page 34: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Discussion

Color and binary histograms do not correlate with the critical view. They do, however, predict when we

are in the abdomen. Currently working on 3D histograms

and spatial-dependency matrices. NIH grant application under

development.

Page 35: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Challenges

Live tissue (vs. solid objects). Deformable. Normal variation. Disease states.

May need to consider. Temporal information. Relevant clinical data of the patient. Critical view "rectangle" (contextual).

Page 36: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Summary

We are comparing image features with similarity metrics to identify the critical view.

This is a first step in automated video summarization, to help with video-assisted evaluation of laparoscopic surgery.

Page 37: Video Summarization of Key Events Stage I - The Critical View Michael A. Grasso, MD, PhD University of Maryland School of Medicine UMBC Computer Science

Thank You