Duke Center for AI in Radiology (DAIR)

Duke Center for Artificial Intelligence in Radiology. Developing AI algorithms and implementing them in clinics

[Maciej A. Mazurowski @MazurowskiPhD]

Does Segment Anything Model work for medical images?

Following our initial analysis a few weeks ago, we conducted thorough experiments on 19 datasets.

The paper is here: https://t.co/pKd0sQI4be

Here is what we learned. 🧵1/9

[Maciej A. Mazurowski @MazurowskiPhD]

Does Segment Anything Model (SAM) really work on medical images?

We tested it on 11 datasets and wrote a paper:

https://t.co/cYaNOSSpRY

In brief: Performance varies widely across different datasets from impressive (given zero-shot setup) to poor.

Details follow. 🧵1/14.

[Duke Center for AI in Radiology (DAIR) @duke_dair]

Check out the results of the machine learning challenge that we co-organized!

[Duke Spark - Medical Imaging AI @duke_spark]

Exciting news! The results of the DBTex Digital Breast Tomosynthesis Lesion Detection Challenge, led by @duke_spark's @nick_konz and @MazurowskiPhD, have just been published in @JAMANetworkOpen! They present new algorithms, a benchmark, and a dataset.
https://t.co/2z4O6JCqkm

[Duke Radiology @DukeRadiology]

Team @DukeRadiology at #RSNA22. @MazurowskiPhD New Modeling Approaches for Radiology AI
E451B M4-CIN21 Nov 28, 11:00 am – 12:00 pm
@duke_dair @DukeAIHealth
@DukeHealth @DukeMedSchool @RSNA

[Duke Radiology @DukeRadiology]

Team @DukeRadiology at #RSNA22. @MazurowskiPhD Multimodal Data Model Development in Radiology: Beyond the Pixel Data E450A S5-CIN20
Sun Nov 27, 2:30 pm – 3:30 pm
@duke_dair @DukeAIHealth
@DukeHealth @DukeMedSchool @RSNA

[Duke Center for AI in Radiology (DAIR) @duke_dair]

Check out this seminar by our scientific director!

[Maciej A. Mazurowski @MazurowskiPhD]

Join us for this seminar. I will talk about our public breast MRI dataset and about public sharing of data in general.

[Maciej A. Mazurowski @MazurowskiPhD]

Join us for this seminar. I will talk about our public breast MRI dataset and about public sharing of data in general.

[Duke Center for AI in Radiology (DAIR) @duke_dair]

@DukeAIHealth @DukeHealth @DukeRadiology @DukeU @Radiology_AI
#AI #medicine #radiology #science #BigData #Analytics #DataScience #AI #MachineLearning #IoT #IIoT #Python #RStats #TensorFlow #Pytorch #DataScientist #Programming #Coding #100DaysofCode (6/6)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

When tested, the automatically extracted labels from the developed rule-based algorithm were 91% to 99% accurate as compared to manual validation set. (4/6)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

All of the model’s weights, code, and hyperparameters are publicly available here. More details and implications can be found in the attached paper.
https://t.co/UZUzSoXl3u
https://t.co/SXk5RCKgwD (5/6)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

The weakly supervised model learned to extract volume level labels through automated rules from radiology text reports to reduce the need for human annotation efforts. (3/6)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

This paper by @duke_dair's @MazurowskiPhD and colleagues combats this issue by utilizing automatically extracted labels from radiology text reports for multi-disease classifiers for body CT scans. (2/6)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

AI algorithms developed for disease classification typically target one disease and organ, limiting its clinical potential. A major challenge for developing multi-disease detection/classification models comes from a shortage of annotated data. (1/6)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

@DukeAIHealth @DukeHealth @DukeRadiology @DukeU @Radiology_AI
#AI #medicine #radiology #science #BigData #Analytics #DataScience #AI #MachineLearning #IoT #IIoT #Python #RStats #TensorFlow #Pytorch #DataScientist #Programming #Coding #100DaysofCode (7/7)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

This paper by @duke_dair's @MazurowskiPhD and colleagues @budamat @ebadawy_ explores whether deep learning can use MRI images to predict lower-grade glioma (LLG) genomic subtypes. (1/7)

[Duke Center for AI in Radiology (DAIR) @duke_dair]

The attached paper offers more detail and implications.
https://t.co/MaLLUgT9Nd (6/7)