71 strong teams at home and abroad competed for glaucoma AI, and the voxel technology team won the first place in the task of cup and plate division

Heart of machine 2021-10-14 02:35:28

MICCAI(Medical Image Computing and Computer Assisted Intervention) Began in 1998 Massachusetts Institute of Technology , Intended to explore medical imaging 、 The value of computer-aided intervention and the integration of the two .20 Over years of development ,MICCAI It has become the top academic conference in the medical image analysis industry .

Seminar on ophthalmic image analysis organized by Baidu OMIA (Ophthalmic Medical Image Analysis) It is one of the key seminars in the field of ophthalmic imaging , So far, it has held eight .2021MICCAI above ,OMIA Focus on glaucoma , Held GAMMA challenge round , Glaucoma artificial intelligence algorithm converging all over the world , Total at home and abroad 71 The branch competes here .

The participants in this competition will be against 300 In pairs 3D The fundus oculi OCT Volume data and 2D Fundus color photo data processing , It includes three sub tasks :1) Glaucoma grade ;2) Macular fovea localization ;3) Optic cup & Optic disc segmentation . In short , Each subtask corresponds to a measurement task AI The key to good and bad .

In this world-class contest , Medical artificial intelligence enterprise voxel technology medical imaging team is established jointly with the imaging Institute of the Department of electronics of Shanghai Jiaotong University Voxelcloud Team Strong , from 71 Stand out from the support brigade , Take the 2 Excellent results of famous , And looking at the cup & The sub task of optic disc segmentation ranked No 1.

chart :GAMMA The final ranking of the challenge

According to the voxel technology team , The reason why we can win the first prize in the competition , Because the model used in the competition gathers a variety of advanced algorithm technologies in the industry , And fully customized by the team .

For the macular localization task , The team adopted the latest in the industry TransUNet[3], And combine U2-Net Some modules segment and locate the macular region . For cup and plate segmentation task , The team came up with a two-stage segmentation strategy : First, make a rough positioning of the cup and plate area , Then, the cup and disc fine segmentation is further performed on the located optic disc area image block .

Besides , The team adopted TransUNet And other advanced network architectures for rough cup and plate area segmentation and positioning , And introduced Polar Transformation[1] To enhance the expression ability of image features . In the fine segmentation stage , The team adopted the latest algorithm framework , Include Segtran[4],TransUnet[3] and CE-Net[2] To split the cup and plate , The average voting method is used to integrate the prediction results of different models to improve the segmentation accuracy .

chart : Stage 1 Rough disc segmentation process

chart : Stage 2 Fine cup and plate segmentation process

Optic cup & Accurate segmentation of optic disc is very important for the evaluation of glaucoma , In clinical practice , Vertical cup to plate ratio (CDR) The bigger it is , The higher the risk of glaucoma . In particular , The optic disc is the part of the retina where the visual fibers converge and pass out of the eye , It's the beginning of the optic nerve . When the optic disc bleeds 、 The rim becomes thinner 、 Vascular knee flexion 、 When the cup plate ratio is enlarged , Glaucoma lesions are likely to occur . therefore , The excellent results of the team in this link lay a good foundation for glaucoma classification and subsequent diagnosis .

Glaucoma is the second blinding eye disease in the world , According to the WHO forecast ,2020 The global primary glaucoma will reach 7964 10, , among 11.2% Our patients will be blind , The total number of primary glaucoma patients in China will reach 2182 10, , Of the world 27.4%. From the data, we can see AI The value of .

Under the massive demand , Voxel technology will continue to study glaucoma AI Auxiliary analysis tools , Help more patients achieve early screening and early treatment of glaucoma , Strangle the hidden dark moments in the cradle .


[1]  Fu, Huazhu, et al. "Joint optic disc and cup segmentation based on multi-label deep network and polar transformation." IEEE transactions on medical imaging 37.7 (2018): 1597-1605.

[2] Gu, Zaiwang, et al. "Ce-net: Context encoder network for 2d medical image segmentation." IEEE transactions on medical imaging 38.10 (2019): 2281-2292.

[3]  Chen, Jieneng, et al. "Transunet: Transformers make strong encoders for medical image segmentation." arXiv preprint arXiv:2102.04306 (2021).

[4] Li, Shaohua, et al. "Medical Image Segmentation using Squeeze-and-Expansion Transformers." arXiv preprint arXiv:2105.09511 (2021).

本文为[Heart of machine]所创,转载请带上原文链接,感谢