Vault

NPB136-L13

3 min read

📗 -> 02/06/26: NPB136-L13

[Lecture Slide Link]

🎤 Vocab

❗ Unit and Larger Context

Why hasn’t AI came for radiologists?

✒️ -> Scratch Notes

Why not?

  • Performance on benchmarks hard to replicate in real world
  • Legal and regulatory issues
  • Radiologists do a lot more than read images

Real world performance:

  • Each model diagnoses one condition, and it’s been hard to build uni ed models
  • Only a fraction of conditions currently have good models
    • If a model is only trained to detect cancer, a radiologist has to look over it anyway to make sure theres no lymphoma. Not much time save
  • Models typically trained on data from one or a few hospitals and don’t generalize. Sometimes reflects differences in ways hospitals take scans.
  • Training data skews towards unambiguous (easy) cases
  • Training data has fewer cases from specic populations. E.g., skin cancer data has many fewer examples with dark skin.
  • Models can do well overall while doing badly on a minority population
  • AI assistance changes human behavior, and can make physicians signicantly worse

Legal and regulatory issues

  • Currently very high bar for fully autonomous systems
  • Assistive tools still require a physician to review and con rm
  • Insurance companies don’t want to provide malpractice insurance. Potential for huge damages—a bad model can harm many many more people than a bad radiologist

Jevons Paradox - Steam engines become more efficient at using coal -> Coal consumption increased

Difference between speed, cost and accuracy

  • Imagine every scan is reviewed by both radiologist and AI. The radiologist takes a second look if the AI disagrees.
  • This procedure should make scans more accurate
  • But will make them slower (since some of them need to be reviewed twice) and most expensive (radiologist reviews fewer scans, AI has to be paid for)
  • Is this a good outcome?
  • Would hospitals pay for this?

In this situation, AI has made radiology more accurate, but also more expensive and slow. Is this a good outcome?

Code for lab of creating neurons

# These functions help set up rectified linear neurons.

def relu(z):
    return max(0, z)

def relu_neuron(input1, input2, weight1, weight2, bias):
    z = weight1*input1 + weight2*input2 + bias
    output = relu(z)
    return(output)


x1, x2 = [0, 0]
w1, w2, w3, w4, w5, w6 = [2, 2, 2, 2, 2, 2]
b1, b2, b3 = [2, 2, 2]

y1 = relu_neuron(x1, x2, w1, w2, b1)
y2 = relu_neuron(x1, x2, w3, w4, b2)
y3 = relu_neuron(y1, y2, w5, w6, b3)
print(y1, y2, y3)

🔗 -> Links

Resources

  • Put useful links here

Connections

  • Link all related words

Graph View

Backlinks