acuity & contrast sensitivity
together define 2D space
—> without both no 2D possible
Acuity
smallest detail you can resolve
contrast sensitivity
faintest difference in luminance you can detect
contrast sensitivity varies with:
spatial frequency
low —> higher contrast sensitivity
mid —> best contrast sensitivity
high —> bad contrast sensitivity
eccentricity
distance of objects from center of gaze
eccentricity effects
acuity decreasesas you move away from fovea
lower density of cones
peripheral crowding
less spatial resolution —> harder perception of individual objects
mid-level vision gestalt grouping rules
proximity elements close to each other: belong together
similarity —> elements similar grouped together
shape
size
color
orientation
closure —> contour completion
continuity —> smooth flowing lines perceived instead of discontinuous or jagged
symmetry —> group or figure
common fate —> move in same direction
radiology types of search errors
search error
recognition error
decision error
UFOV
useful field of view
area of visual space within which information can be perceived and processed without eye or head movements
Search error Kundel
target never fell into UFOV
targets that fell into UFOV for <1 second
if target attracted foveal vision for >1 second
then: decision to deem irrelevant
difference 2D and volumetric search (3D)
volumetric: hundreds or thousands of images
all roi cannot be foveated exhaustively (rely on peripheral vision -> low accuracy)
in CT —> only between 27 and 69% of lung tissue foveated
but: dynamic (motion-based) information is often used
—> observer performance in 2D search might not be generalizable to volumetric tasks
search path in 3D volumetric space
two strategy types
Scanners
start at top
left to right
move down
Drillers
eye in one quadrant and move up and down
Drillers outperform scanners in detection
Holistic gist processing theory
scenes can be understood very quickly (“gist”)
but details & objects are forgotten
—> could there be a gist of breast cancer
—> radiologists have a “feeling” -> subconscious
___
flash mammogram for 250 msec —> would they call patient back?
outperformed chance
UNEXPLOITED IMAGING BIOMARKER
Dual pathway model for visual processing
for naturalistic image processing
can be adapted to radiology
selective pathway
non-selective pathway
dual pathway model - selective pathway
often requires attention
binds features and recognizes objects
capacity-limited —> represented as bottleneck
controlled by guidance mechanisms from non-selective pathway
prioritizes likely target items
classic guidance prefers items with basic target features e.g. color
Non-selective pathway
guides selective pathway via important features (color, orientation etc.)
semantic elements
episodic elements
no capacity limitation
extracts global statistics from entire scene
enables certain amount of semantic processing
no precise object recognition but scene categorization
difference novice radiologists vs. expert radiologists
experts:
know where not to look
faster at fixation on abnormalities
total image search time decreases with increasing level of expertise
fewer total fixations than novices
—> more effective
—> perceptual aspects of image interpretation develop earlier than correct interpretation of abnormalities or integration into diagnosis
—> perceptual skills grow from start of exposure
—> radiology-specific factual knowledge contributes little to this initial development
Perceptual expertise - machine learning
machines still make errors (muffin/chihuahua challenge)
—> radiologists are encouraged to work with AI as collective intelligence
—> radiology is perceptual task-dominated field
—> radiology needs greater understanding of perceptual expertise to improve accuracy, reduce error
geoffrey hinton - radiology and machine learning
radiology as specialty will become obsolete due to exponential improvement oft deep learning medical image analysis software —> not true up until now
Obstacles for radiology AI
no standardized training datasets
poor interoperability standards
regional variances in image acquisition/processing
delayed adoption of technological advances in healthcare industry
Augmented intelligence in radiology
radiologists could capitalize on clinical judgments about data not detection
AI could train radiologists —> provide them with images they tend to make most mistakes on
Perceptual expertise definition
is fundamental to most activities
computer system often make major errors in comparison to humans
example of perceptual expertise: radiologist
Last changed2 months ago