Alibaba Open-Sourced Model That Detects Cancer From CT Scans - I Ran It Locally

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Alibaba Open-Sourced Model That Detects Cancer From CT Scans - I Ran It Locally · Channel: Fahd Mirza · 07:52 · Watch on YouTube

Topics: Local and open-source models, Health and science

Summary

Fahad Mirza introduces RADAR, an open-source Alibaba model designed to analyze abdominal CT scans and identify nearly 150 conditions, including several cancers. He describes its training data and reported research results, then demonstrates installing the software, downloading its model and organ-mask data, and running inference locally. The output is a CSV of confidence scores for 146 conditions, which the creator says should be reviewed by qualified medical professionals rather than treated as a diagnosis. He highlights the model’s potential to support radiologists in underserved and remote areas.

Key points

RADAR was built by Alibaba’s Damo Academy to analyze abdominal CT scans for nearly 150 conditions.
The video says the model was trained on over 400,000 CT exams and 15 million image-report pairs, without manual annotation.
The transcript reports an AUC of 0.913 in a study of 40,000 real-world exams, and says RADAR outperformed 23 of 26 radiologists in a reader study while boosting their sensitivity by 10%.
The creator installs the repository requirements and downloads the pretrained model and organ-mask data before running the inference script.
Organ masks provide templates indicating where organs such as the liver, pancreas, and stomach are located in the scan.
The demonstration produces a CSV with one row per case and 146 confidence-score columns.
The example output includes aortic calcification at 0.70 and sclerosis at 0.58, which the creator says are high enough to flag for radiologist review.
The creator emphasizes that he is not a medical specialist and frames the model as a potential aid for qualified clinicians, particularly where radiology services are limited.
He notes that the model can run locally on a GPU, though reported VRAM use rises during the process, and mentions script settings that may reduce memory requirements.

Chapters

Start	Chapter	Summary
00:01	RADAR and its potential	The creator introduces Alibaba’s open-source medical model and explains how local access could help hospitals with limited resources.
01:07	Installing the model and reviewing its claims	He begins setting up the repository and outlines RADAR’s intended use, training data, and reported evaluation results.
02:41	Downloading data and explaining the workflow	The demonstration downloads the model and organ masks while the creator explains how scans, reports, and organ regions fit into RADAR’s workflow.
04:16	Running inference and reading the output	He runs the inference script, discusses its GPU memory use, and interprets the resulting confidence-score CSV as material for radiologist review.
06:19	Clinical limitations and access	The creator cautions that he is not a medical specialist and discusses how the model might support care in remote communities.

Quotes

Alibaba open sources medical AI model that can detect cancer and nearly 150 conditions. — 00:01

It was trained on over 400,000 CT exams with 15 million image report pairs. — 01:37

the model then should be outputting a scoring for each of the 146 conditions. — 04:47

I’m not going to pretend that I’m a medical specialist — 06:19

Really huge kudos to Alibaba because not only they have released it, the license is quite liberal, too. — 07:21

Concepts and entities

RADAR, Alibaba, Damo Academy, Hugging Face, Mass Compute, Fahd Mirza, Máscara de órgãos, Pontuação de confiança

Full transcript (original language)

00:01 Something really noble and cool happened this weekend. Alibaba open sources medical AI model that can detect cancer and nearly 150 conditions. Cancer screening does more for AI’s public reputation than any chatbot demo. Open weights mean the hospitals that cannot afford an API gets to run it anyway. Take a rural district hospital, let’s say in Bangladesh or Sri Lanka, or a clinic in sub-Saharan Africa. No

00:33 radiologist on staff, no budget for a per scan diagnostic API, but a decent GPU can now run this model for free. This is Fahad Mirza and I welcome you to the channel. In this video, we are going to install this model and we are going to test it out. Recently, I also have started this free weekly AI newsletter, which you can subscribe at fahadmirza.substack.com. And you know what? More than 40% of my viewership is either from um Asia or

01:07 Africa and I’m so glad that this model would be so helpful, especially in those countries, or maybe across the globe. So, let’s get it installed and we will talk more around what exactly this model is from the architectural perspective. I’m going to get clone this repo and I will drop the link in video’s description. It’s a big repo, by the way. It’s not a small repo. So, while it happens, I think it is going to take a minute. That’s done. Let me quickly install

01:37 all the requirements from the root of the repo. And while that happens, let’s talk a bit more around this model. So, this model is called as Radar. Built by Alibaba’s Damo Academy, it reads abdominal abdominal CT scans and identifies nearly 150 conditions, including liver, pancreas, stomach, and colorectal cancers. It was trained on over 400,000 CT exams with 15 million image report pairs. No manual annotation

02:09 required. In a science study across 40,000 real-world exams, it hit an AUC of 0.913. And in a reader study, it outperformed 23 of 26 radiologists while boosting their sensitivity by 10%. So, that is why I decided to cover it in this model. There’s a lot of information which they have shared in this science.org report. And then the model code also list down a GitHub repo and some other code which I’m going to use in this video. I’m also

02:41 going to explain what this image means, but for now, let’s get it installed and see how it works. If you’re looking to rent a GPU on very good price, you can find the link to Mass Compute in video’s description with a discount coupon code of 50% for a range of GPUs. Okay, let’s go back to terminal. Everything is now installed, as you can see. Next up, let’s download the model and data. And this is a script which I’m going to use to download both.

03:12 And it is going to take a bit of a time. And by the way, the data which I’m downloading here, as you can see, the second script, it is a pre-made outlines of organs, like a template showing where the liver, pancreas, stomach, etc. sit in a CT scan. RADAR uses these to know which part of the scan to look at before it diagnoses anything. While it runs, let’s have a very quick look at this diagram and what it means. So, primarily, this diagram is showing how RADAR works end-to-end. On the left,

03:44 thousands of CT scans get paired with their radiology reports and mashed up through vision language alignment, teaching the model to connect what it sees with what doctors wrote. In the middle, once trained, readers can see abdomen and flags problem areas across organs like the liver, pancreas, and stomach. On the right, it acts as a second pair of eyes for radiologists, boosting their diagnostic sensitivity while they read the scan. So, that is all there is to it.

04:16 And you can see that all the data is being downloaded, so let’s wait. One thing to note here is that you can see that it rate limited me because I’m using free hugging face account. You know what? No need to buy anything from them. Just go and rerun it. It downloads everything again. So, everything is done. The good thing is that in their repo, they have also shared an inference script, which I’m going to run now. So, you need to be in radar_inference directory. This is going to load our

04:47 pre-trained checkpoint and the bird text model. It is going to feed in a real abdominal CT scan and its organ mask and the model then should be outputting a scoring for each of the 146 conditions. So, let me run this. And you can ignore these warnings if you get them cuz they are just about packages and stuff. It is initializing the model, loading, and you can see just under 1 gig of VRAM. So, you can easily run it on um a commodity GPU.

05:18 So, it is it has jumped to around 16 gig, but still you can run it in under commodity GPU easily. There is also uh where you can reduce some of the parameters like KV cache and stuff, context length, in the script, which is going to even reduce it furthermore. But anyway, um if you look at this output, this is quite interesting. This is the actual output it generates while working through it after reading the scan. And if I scroll through, this is a CSV file

05:49 it has generated and it looks something like this. So, one row per case 146 columns of confidence scores. The file name identifies which CT case this is and each column headed name as an organ paired with a specific condition it’s checking for. Every number is the model’s confidence that condition is present. So, higher means stronger evidence in the scan. I have just taken a bit of a sample here as you can see. So, this is a trimmed sample from that

06:19 exact file showing just the aorta findings for this case. So, which is calcification score 0.70 and some other stuff like which is sclerosis 0.58 both high enough to flag for radiologist review. I have no idea what exactly this mean. I’m not going to pretend that I’m a medical specialist, but I’m more than sure this is quite useful if you are in that field and of course for a trained qualified physician, radiologist, or whoever is

06:50 the responsible person I think this is quite useful and as I said the main use case is for the house of hospital radiologist who are working in the trenches in remote places or in places where there are no facilities and that is even true for the first world countries. For example, here in Outback Australia on remote communities there is hardly any proper facility. You would be surprised to learn that they have to either fly or travel thousands

07:21 of kilometers to get a proper medical attention. And I think these sort of models really make it uh quite possible for medical services to reach to the masses quite easily. So, that’s it. Let me know what do you think about this model. Really huge kudos to Alibaba because not only they have released it, the license is quite liberal, too. There are, I think, here and there a bit of a uh thing which you can argue about, but

07:52 it is still quite liberal. And all the files, all the data set, everything is present in their hugging face card, which I will drop the link in video’s description. That’s it. If you want to help out the channel, please become a member. Thank you for all the support.