← Lectures·L21

AI in Communication Disorders / Resource-Efficient AI

Kist · 1 concepts · 7 questions

Concept 1 / 1

Key Concepts to Memorize

ANKI Lab = Artificial Intelligence in Communication Disorders


  • Applies AI to disorders affecting communication (speech, swallowing)

Central Dogma of Neuroscience / Kist's Framework:


Sensory Input → Brain (Processing) → Behavior (Output)

  • AI can intervene at any stage: sensing, processing, or behavior analysis

Dysphagia (Swallowing Disorder):


  • Dysphagia = difficulty swallowing (Greek: dys = difficult, phagein = to eat)
  • Can lead to aspiration (food/liquid enters lungs) → pneumonia, malnutrition
  • Analyzed using fluoroscopy (X-ray video) — the clinical gold standard
  • AI challenge: automatic, objective analysis of swallowing from video

Key AI Technique — Semantic Segmentation:


  • Semantic segmentation = pixel-by-pixel classification of an image
  • Every pixel is assigned a class label (e.g., "tongue", "food bolus", "airway")
  • More detailed than object detection (which only gives bounding boxes)

Markerless Tracking:


  • Track structures (e.g., tongue, hyoid bone) in video WITHOUT physical markers
  • Method: compute velocity fields → integrate over time → displacement fields
  • Enables tracking anatomical movement without surgical markers or radiation markers

Diffeomorphic Image Registration:


  • A method to align (register) two images
  • Diffeomorphic = smooth, invertible transformation (no tearing or folding of tissue)
  • Preserves anatomical topology

Developmental Speech Disorders:


  • Affect 10–20% of children
  • Two main types:

- Sensory type: Cannot hear/perceive their own speech correctly (auditory feedback problem)

- Motor type: Knows what to say but has difficulty with execution (motor coordination problem)


Phoneme — Definition:


  • Phoneme = the smallest unit of sound in a language that distinguishes meaning
  • Example: /p/ vs /b/ in "pat" vs "bat"

TTS — Text-to-Speech:


Text → AI Model → Audio output

  • AI generates natural-sounding speech from written text
  • Application: assistive technology for people who cannot speak

Embedded AI (Resource-Efficient AI):


  • Running AI models directly on the device (not sending data to the cloud)
  • Constraints: limited memory, limited compute, limited energy (battery-powered devices)
  • Same concept as Edge AI / TinyML
  • Key trade-off: model accuracy vs. model size/energy cost