Images (Introduction to Medical Informatics) (http://www.cpmc.columbia.edu/edu/textbook) LAST REVIEWED: 1 October 1997 IMAGES a major source is the radiology department image generation (define each) film-based X-rays: produces a "shadowgraph" that depicts a directly measurable parameter (X-ray absorption) digital radiology: DSA and digitized images computer tomography ultrasound imaging nuclear medicine imaging magnetic resonance imaging image analysis visual: determine if features of disease present quantify: volume or mass localization of lesions screening for disease interpretation recording: visual analysis transcription of dictation bar code or enunciation of standard report code, with patient-specific customization structured templates (e.g., UltraSTAR at BWH) speech recognition software image management film libraries are expensive and inefficient film can only be in one place picture archiving and communications system (PACS) digital image library how to get the image start with film and take picture go directly to digital form via phosphor media some are inherently computerized then store in a database transfer over network view on display archive old images (need to keep old headers) image properties relating to storage pixel = smallest indivisible picture element voxel = element of a 3-D volume spatial resolution (sharpness): ability to distinguish two points = # pixels / image area line pair / mm cycle / degree of vision (about 6 is ideal, 50 is max, CRT is 12) contrast resolution: ability to distinguish differences in intensity = # bits / pixel temporal resolution: time needed to create an image = shutter speed (30/sec for heart) real-time = > 30 images/sec eg, place lines on board in 1/8" cycles (1/16" width) front of room = 10 ft => 16 cycles/degree (visible) 6*120inches=720in circumference 8 cycle/in*720in = 5760 cycle/circle 5760/360degree = 16 cycles/degree back of room = 30 ft => 48 cycles/degree (barely seen) storage capacity CT: each pixel is derived from multiple measurements 512 x 512 pixels 11-13 bits / pixel 11-30 slices 100 million bits (12.5 megabytes) chest X-ray: each pixel is a direct measurement 2048 x 2048 pixels 2.5 cycles per mm (512 => 0.6 line pairs) 14 bits / pixel 2 views 110 million bits (13 megabytes) (vs. paper which was 512x480@8bit=250,000bytes) other studies are smaller data compression x3 with no loss x8 with some loss department 250 exams / day 25 million bits / study (3 MB) 250 days /year 10^12 bits / year = 100 GB 100 optical disks = 10^12 bits, costs $10,000 one study of all radiology in 500 bed hospital showed 900 MB/day => 250 GB/year (r/e archiving) communications 10^4 bits per second (10 Kbps) => 3 hours per chest X-ray (vs. 3 minutes for paper with 512, compression) 10^7 bits per second (10 Mbps) => 10 seconds per chest X-ray but not actually move data so quickly 2048 is 16 times as much data as 512 therefore, need high speed networks in first three days of hospitalization => 10 film requests then 4 over next 6 days then 3 over next year calculated need 25 megabit per second network workstation design and expense 2048 x 2048 is much more $ than 1024 x 1024 can use lower resolution and zoom tradeoffs resolution/price/storage space/transmission time vs diagnostic accuracy using film to diagnose is more accurate than digital image film is equivalent of 2000-4000 as hardware improves, digital approaches film clinical trials: 1. 59 radiographs by six radiologists compared to 3 others on film 512x480, 8 bit selected subtle findings like difficult fracture res %err 512 50% 1024 29% 2048 1.4% 2. 919 cases shipped from 535 urban hosp to 780 bed hosp 3 miles away 512x512 images, 8 bit resolution 2.5 compression, 9,600 baud modem, 3 minutes per image read at night, then compared to final reading included normals 22 were lost (? due to paper system) significant discrepancies due to image in 1.6% (14) of 897 cases and 4.3% due to reader error or reader variability image problems related most problem on pneumothorax and abdominal calcification 3. panel of radiologists 512x512, 8 bit included normals 1.9% of 4028 interpretations were errors digital had higher sens, lower spec, lower accuracy repeat with 1024x1024 => not better image processing (advantage of computers) width/level = "good" histogram equalization (gray scale augmentation) = "fair" reverse video = "fair" zoom (in paper, no gain in resolution) = "not use" reversal rotation edge enhancement and smoothing mark area for later review = "not use" CPMC example radiology images from distant Allen Pav. to Milstein to save radiologist FTE fell out of use: difficult scanning, slow transmission only inherently computerized images (CT) worked well additional DB requirements (other than storage space) performance separate headers key words thumbnail sketches interpretation of images human error 20% on detection 10-50% on diagnosis miss signs of disease embedded in noise automated image analysis computer processes image eg, heart ejection fraction has not worked well now focus is enhancing images eg, 3D image generation for CT, MRI related reading: Markivee CR, Mahanta B, Savci S, Abbas S, Chenoweth JL, Luther R, et. al. Diagnostic accuracy of a teleradiology image transmission system. M.D. Computing 1989;6(2):88-93.