The medical version of SAM is open source! A universal segmentation model for medical images is here!

The medical imaging industry has always built very high field barriers based on data and high-cost annotation. With the recent introduction of general CV large models such as SAM and SegGPT, large-scale Models and general models are also gradually burning in the field of CV, especially in the field of image segmentation. Various general medical image segmentation models have also emerged .

On 11.16-11.17, we invited Ph.D. from Taiwan Chiao Tung University, author of many top articles, Mr. Shawn< a i=3>, brings us - New SOTA for general cross-modal medical image segmentation, and explains the medical image segmentation task in detail.

Scan the QR code to make an appointment for the live broadcast (free PPT for teaching by the teacher)

Free collection of MICCAI2023 popular papers recommended by leadership teachers

89ec406b5359406b41fe8dff07ca0fc3.png

195b5fc628839cdc55069140ef20669f.png

MICCAI2023 popular medical imaging paper collection

Teacher introduction: Teacher Shawn

-Ph.D., Taiwan Chiao Tung University

-Published multiple papers as the first author, including ICLR, ICDE

-Won multiple school-level scholarships, AI competitions, and cooperated with the Singapore Ministry of Science and Technology

-Research directions: deep learning, computer vision, music generation, multi-modality

Live broadcast outline

1) For medical image segmentation, if a model is trained using only MR images in the source domain, how good is its performance in directly segmenting CT images in the target domain?

2) For generalized medical image segmentation tasks, it is very difficult to train a model using a single source domain. How to solve this situation?

Scan the QR code to make an appointment for the live broadcast (free PPT for teaching by the teacher)

Free collection of MICCAI2023 popular papers recommended by leadership teachers

793d8470bbd4062643a2a7a806c6983a.png

97867ce5594d402662f4be698e2641c3.jpeg

For medical image segmentation, imagine if a model is trained using only MR images in the source domain, how good is its performance in directly segmenting CT images in the target domain? This setup, i.e., generalizable cross-modal segmentation, has clinical potential , more challenging than other related settings such as domain adaptation.

For generalized medical image segmentation tasks, it is very difficult to train a model using a single source domain. Stylistic deviations between different modes can significantly degrade performance. As shown in the figure below, different modalities of brain tumor impact have significantly different appearances, as shown in the figure below, T1 and T2. Existing models often only perform well in a single modality and perform poorly in another modality, such as DeepAll and DoFE methods.

c4d3757bb13b6e93c9e184a43285703d.png

Recently, some research has proposedSAM-Med3D, which is a three-dimensional SAM model specially used for 3D voxel medical image segmentation. . This model significantly outperforms other methods while providing limited cue points for different anatomical structures such as bones, heart, and muscles. Under different image modalities, especially MRI images, more cue points are usually required to achieve the same performance than CT images, but SAM-Med3D performs well under various modalities (including MRI images), organs, and lesions. Always perform well. In addition, the transferability of SAM-Med3D has also been verified on different benchmark tasks, and the model has shown strong potential, so SAM-Med3D is expected to become a powerful pre-training model for 3D medical image Transformer.

b190de15835db1087fdc2fb891bd122c.png

SAM-Med3D with complete 3D structure

6a0fcdf3fa298c2767b68f5858bf3cc2.png

Scan the QR code to make an appointment for the live broadcast (free PPT for teaching by the teacher)

Free collection of MICCAI2023 popular papers recommended by leadership teachers

ef604b67d6addf9f11778292bc0fd5d9.png

As a novice in scientific research, how can I publish a high-quality paper?

For the paper, everyone is working hard to design new networks, new strategies, and new training algorithms. As long as we can achieve a good performance on a certain problem, the paper will be completed naturally. And if you want to achieve it quickly, guidance from seniors is indispensable.

The role of a good instructor is that if there is no topic, he can help you plan the topic based on the specific situation of the topic group and recent popular research directions. If there is a topic but lacks innovative directions, the teacher can quickly help you find several entry points. , several frameworks, and even help you think about which documents you need to read...

Scan QR code

One-on-one meeting with Daniel mentor

574cce3af2ada1b372618d6fcebc93c8.png

6dcec99e03c05531e4e4051fb636bc9c.png

MICCAI2023 popular medical imaging paper collection

Welfare at the end of the article

Xiaowo has compiled a series of premium paid courses from Mr. Paul, co-founder of Vaughn Wisdom, The original price is 3999 yuan. Now you can get it for 0 yuan for free. It includes hot topics in all aspects of the computer field and essay writing tips!

b47df61dc79e00bbb75e21e23f828730.png

d2494fed9fc9d71e70df0fcca6101512.png

8c7f574a6d0266ce77aacc753290f914.png

Scan the code now to receive a series of courses for free

-END-

Guess you like

Origin blog.csdn.net/woshicver/article/details/134453720