Week Eleven - Computer-Aided Diagnosis: From Macro to Micro Flashcards
What does CAD stand for?
Computer aided diagnosis
What is CAD?
A diagnosis that is made by a physician whom is assisted by the output of a computerised analysis of a medical image.
What fields of research does CAD involve?
- Medical imaging
- Digital image processing
- Pattern recognition/machine learning/artificial intelligence.
What are the three main things CAD can do?
- Increase efficiency and effectiveness of diagnosis by acting as a ‘second opinion’ for physicians.
- Extract and analyse the characteristics of benign and malignant tumours.
- Objectively and quantitatively classify patterns as diseases.
Why do we need CAD? (2 reasons)
- Medical images are currently analysed through visual inspection on a slice by slice basis; which requires a high degree of skill, is expensive, time consuming, prone to operator bias, and unsuitable for the processing of large scale research samples.
- Rapid increase of patients and only a small increase of physicians.
Describe CAD in the 1960s.
Automated computer diagnosis (ACD) was attempted bc they thought computers were better at performing certain tasks than humans.
Why did CAD fail in the 1960s?
- Computers not powerful enough.
- Advanced image processing techniques weren’t available
- Digital images were not easily accessible.
What were three major research topics in the 1980s?
- Detection of lesions involved in vascular imaging.
- Detection of lung nodules in chest radiographs.
- Detection of clustered microcalcifications in mammograms.
What would the ultimate success of CAD look like?
Daily use in routine clinical work at hospitals worldwide.
Name the three types of CAD systems.
CADe - detection of lesions.
CADq - quantifications (measure size of tumour etc)
CADx - diagnosis
Name three applications of CAD.
Mammograms.
Chest Radiography.
Histopathology.
Describe the use of CAD in mammography.
15-30% of clustered microcalcifications are missed! FDA approved commercial system in 1997 which identified 80% of missed lesions in one study.
Describe the use of CAD in chest radiography.
Radiologists miss 30% of lung nodules bc of the camouflaging effect of the normal anatomic background.
Why hasn’t CAD become common in pathology? (2 reasons)
- Histopathology images are very large (giga-pixels)
2. Difficult to localise small tumour regions.
What are two requirements for a good CAD system?
- Sufficient sensitivity and specificity.
2. Good U.I.
What are the 5 steps to building a CAD system?
- Collect relevant normal/pathological cases.
- Develop computer algorithm.
- Validate computer algorithm.
- Evaluate radiologists in diagnosis with and without system.
- Clinical trials.
Name 5 methods of feature extraction.
Gabor, LBP, HOG, SIFT, SC
What is a Gabor filter?
A linear filter used for textural analysis.
What does LBP stand for and what is it?
Local binary patterns are a texture operator which labels the pixels on an image by thresholding the neighbourhood of each pixel and assigning the result a binary number.
What is HOG?
Histogram of object gradients - main premise is that an object can be represented by the distribution of intensity gradients. Used for general object recognition such as people or cars.
What is SIFT?
Scale invariant feature transform. Robust against scale, rotation, illumination and viewpoint.
What is SC?
Shape context describes the shape of an object for measuring similarity across images.
Name 4 classification methods.
k-means, k-NN, SVM, SR
How does k-Means classification work?
Discovers clusters of similar values by continually updating the cluster centre as the mean of the entire cluster.
How does k-NN work?
K-Nearest Neighbours, an object is classified by a majority vote of its neighbours, with the object being assigned to the class most common among its k nearest neighbours.
What is SVM?
Support Vector Machine - a classifier that outputs the optimal hyperplane which catagorises data.
What is SR?
Sparse representation - a sparse vector that approximately solves a system of equations.
Name 5 ways to evaluate performance?
Accuracy, sensitivity, specificity, ROC curve, robustness.
What is accuracy in terms of CAD?
Ability to differentiate between healthy and patients.
What is sensitivity in terms of CAD?
Ability to determine the patient cases correctly.
What is specificity in terms of CAD? What does bad specificity mean?
Ability to determine healthy cases correctly. Bad specificity means a high false positive rate.
What is a ROC curve?
Receiver operating characteristic curve. A graphical plot that illustrates the diagnostic ability of a classifier system as its discrimination threshold is varied.
What does a ROC curve plot?
True positive rate (sensitivity) against false positive rate (1 - specificity)
Why is robustness important?
We cannot assume that a CAD system that works with one database will work with another.
What does the future of CAD look like?
CAD assembled as packages and implemented as part of PACs.
Used as a tool for lesion detection and diagnosis.
