Salient facts about me. I

- recently completely Board certification in Radiology!

- am presently doing a fellowship year at UPMC that is 50% Thoracic Imaging, 50% Radiology Informatics

Presently looking at issues in radiology reporting:
• In the old days, stacks of films would arrive in the reading rooms with "requisition forms" paper clipped to each study. Each requisition form would have a barcode printed that uniquely identified the study in the RIS (radiology information system). The films would be hung on a viewbox, and the study dictated by swiping barcode reader attached to a telephone-based dictation system, then dictating into the handset. Using a barcode instead of pressing digits reduced manual data entry time and errors. Exams would be human transcribed within 24 hrs, and signed/FAXed within 48 hrs.
• At present, we have "PACS" (Picture Archive and Communication Systems) that allow instant electronic distribution of radiologic studies to both clinician offices and radiology reading rooms. Furthermore, we can funnel data from physically remote outpatient imaging centers to central reading rooms. As a result, volume of studies being read is increasing. At the same time, clinicians are now getting the study images in their offices within moments after the patient leaves the scanner. They are no longer willing to wait 48 hours for a report, and are increasingly pressing for same-day service - patient gets a scan, then goes to clinician office - they want a final report by the time of a clinical appointment.
• Presently investigating ways of streamlining workflow to facilitate report creation / distribution, looking at what types of context are available in the daily workflow that can be used to facilitate report creation, i.e. intelligent presentation of a list of possible structured reporting template, then just fill out the words, vocabulary / language model selection for recognition.
• We have a lot of contextual information about what is going to be dictated (Modality [CT, MR, "plain film", US, etc], body part, ordering physician / referring physician names) on a case-by-case basis, and this could allow us to be a lot smarter in our choice of language model and dictionary used for speech recognition.
• Most commercial packages focus on realtime recognition, have a single large "Radiology" or even "Medical" vocabulary and language model, and don't take advantage of the additional context available in the workflow. Non-realtime "batch mode" recognition could still be very useful and might allow greater accuracy.

- completed Residency in Diagnostic Radiology at UPMC ('98-'02)

- was formerly an intern at Sacred Heart Medical Center in Spokane, WA.

• internship year there spent in their Transitional Program ('97-'98)
• View of the hospital from my apartment (JPEG 151K)
• View from the roof of the hospital (JPEG 90K)
• link to the Sacred Heart Golf Tournament
• link to Sacred Heart Transitional Program graduation

- am an MD/PhD graduate

• University of Pittsburgh School of Medicine
  
  • Graduated May 1997
• Carnegie Mellon University, ECE PhD graduate 5/19/96
  
  Research interests included (still include, just no time this year!):
  • Portable computing Enthralled with the potential of the Palm and similar platforms to put large amounts of frequently-updated data at a users fingertips, especially medical data such as patient issue lists, drug information, and toxicology information. This is my 8MB hack for the Palm m100.
  • Medical imaging. Significant interests in various tomographic imaging techniques, especially using algorithms which parallelize well.
  • A closely related area of interest is Machine vision, in particular low-latency multi-baseline stereo imaging.
  • The enabling technology for both of these was Parallel Computing, and this was where I actually did my thesis work.
  • Actually, my work dealt with several small subsets of Parallel Computing, namely Barrier Synchonization and realtime scheduling.

    Parallel computing is a simple concept: Take a big task, split it into a set of smaller tasks, and give each task to a separate computer. But, making it work in a scalable, reliable manner is difficult.

    As the number of processors increases, the latency of the communication between processors becomes a limiting factor. In fact, scheduling scarce communications resources for even just a 64-processor machine is like choreographing a performance of ballet dancers and machine-gunners. If everything works just right, one sees an awesome display of numerical grace and firepower. If something goes wrong, though, things get so bad so fast that it takes two weeks just to figure out what happened in that fatal split-second. Debugging parallel programs on machines which are generating over a billion wrong numbers per second is a considerable challenge and a sure-fire way to blow a weekend.

    Full text of my thesis "Tasks and Connection Sets: Choreographed Communication on a Connection-based Parallel Computer" is available in Adobe Acrobat format or Postscript format.

- operate Amateur radio - licensed as KA8HML

• former member of W3VC, the Carnegie Tech (Amateur) Radio Club

- avidly enjoy "Ask Dr. Science" / Duck's Breath Mystery Theatre

A rather long (and slightly outdated) biographical sketch describing who I was, and what the heck I was trying to do in Pittsburgh