Blog Logo

5 Jul 2025 ~ 2 min read

What does cancer look like?


Disclaimer

To be clear, there is NO SINGLE WAY to describe all cancer from any given stand point, otherwise radiology should be solved by now 1

It can take many form, and here are just a few examples: So, claiming I can give your a wordy way to teach you HOW to spot cancer from medical images would be a HUGE OVERSTATEMENT Instead, my goal here is to help you understand WHY cancer look that way, and this builds intuition of HOW to build a system that spot cancer

What is cancer

Cancer is a group of diseases involving abnormal cell growth with the potential to invade or spread to other parts of the body 1 the simple description from wikipedia actually contain 3 key characteristics

  1. grow bigger (rapidly)
  2. invade surrounding organ
  3. move (replicate) elsewhere

and your body are “corrupted” in few ways

  1. it becomes hard due to rapid growth and fighting with surrounding (Desmoplasia)
  2. promoting blood vessel for nutrients supply
  3. surrounding lynph nodes are infected

The tell tale signs

the above characteristics shall be reflected in medical imaging, though may have different appearence across different modality x body region x patient

  1. have in conspicuous “object”; especially true if you have followup scans (longitudinal imaging studies) and really measure the growth
  2. the boundary becomes obscured , spiculated
  3. more than single occurance
  4. this is somewhat subtle and varies alot from different modality, but the general theme here is “inhomogenity” and patterned calcifications
  5. irregular vessel pattern / blood flow (in Ultrasound you can visually see the blood flow into the foreign blob)
  6. looking for surrounding lynph nodes that are accumulated, enlarged, distorted in shape and texture

build a algorithm

  1. you then start building a intrumental “cancer” detector looking for

References

  1. TODO: Hinton’s prediction link

Hi, I'm Qianyi. I'm a ML engineer based in Beijing. read more about me on my website.