
MD, PGY-3 Radiology @BWHRadEdu Kaggle Grandmaster, healthcare + AI Views my own.
Boston, MA
Joined April 2020
- Tweets56
- Following57
- Followers1.3K
- Likes156
Pinned Tweet
The medals have been awarded, and it's official! I finally became a Kaggle Grandmaster, which I never thought would happen.
I have to thank @FelipeKitamura, @alexandrecadrin, and @DrJHoward for carrying me to my first 4 medals.
Pediatric bone age model (G&P) now on @huggingface Spaces: huggingface.co/spaces/ianpan…
Trained on the @RSNA 2017 challenge data, MAE 4.45 on the test set.
Demo purposes only. I am not responsible for how you use the output of this model. 🤗
Excited to have finished in the top 10 in this year's @RSNA @TheASNR @The_ASSR C-spine Fracture Detection AI Challenge on @kaggle!
Congrats to all of the other winners & competitors, and big thanks to the organizers and annotators. Solution summary here: tinyurl.com/4zayhwnp
Ian Pan retweeted
Mon 11/29, 4:30pm & Wed 12/1, 9:30am:
Oooh, a double-feature! Not one, but 2⃣ chances to have @kaggle winners @FelipeKitamura from @UNIFESP_Rad & grandmaster @ianpanmd from @BWHRadEdu show you how they won the pneumonia detection challenge.
#RSNA21
thelancet.com/journals/landi…
Radiologists: 0.713 AUC
Radiologists+AI: 0.808 AUC!
AI alone: 0.957 AUC ...?
Should I take back my previous tweet? 😂
@DrLukeOR
Most of my success is 2/2 open access medical imaging. I'm super grateful to those who've made data sharing a top priority.
In particular, SO to @StanfordAIMI @StanfordRad who've proven time and time again that HIPAA and academic bureaucracy are not insurmountable obstacles.
Also @unifesp and @FelipeKitamura who've also provided data for the past 2 RSNA challenges. I know Felipe loves to Kaggle but chooses to share data from his institution even if it means he can't compete.
1
22
If anyone wants to listen to me ramble for 30 minutes 😂
In the latest episode of the @Radiology_AI podcast, @PaulYiMD interviews @ianpanmd on machine learning, data science and becoming the first physician @Kaggle Grandmaster. https://xitter.gq/t.co/GxzblQGqWB #RadAI #RadPodcasts
Results are official! I'm happy to have won 2nd place in the @RSNA @thoracicrad PE Detection challenge on @kaggle. I've also sneaked into the top 10 for competitions.
Read more about my solution here:
kaggle.com/c/rsna-str-pulmon…
99% sensitive, 99% specific model for finding with 10% prevalence has PPV of 91.7% = ~1 FP/12 predicted +.
5% prevalence, PPV 83.9% = ~1 FP/6 predicted +.
5% prev, 95% sens, 95% spec, PPV 50.0% = 1 FP/2 predicted +.
Pre-test probability is important.
One issue with separate models for each minority group is that now you have to assign patients to groups and use the corresponding model. A lot of bias and assumptions go into assigning patients into groups.
"Prevalent schemes of multiethnic machine learning are prone to generating significant model performance disparities between ethnic groups. These disparities are caused by data inequality and data distribution discrepancies between ethnic groups."
nature.com/articles/s41467-0…
Agree 100%. I hate saliency maps.
Replying to @quantrad
Trying to interpret saliency maps is asking for trouble. The example I like to use is detecting a brain tumor via midline shift (or vice versa).
My go-to recommendation for classification interpretability is to actually train a segmentation model and use that to classify.
Just because I'm a radiology prelim doesn't mean I know how to read this MRI abdomen that we didn't really need to get
1
8
Publishing a dataset is way more valuable than publishing a paper about how you trained a model on a dataset.
Radiology AI is nice. Charge patients for imaging, then keep their data. Sell that data to companies or use it to build your own models. Then, charge them again when you use these models to read their imaging.
Someone once told me to stop wasting my time on Kaggle because it has no academic value.
My answer: I don't care. I like Kaggle. It's fun (way more fun than research). I learn a lot. That's why I do it.







