Multi-label classification and multi-label segmentation of medical data, binary classification of medical data
write in front
I participated in two projects during the summer vacation and gained a lot. There are many detours and experiences in building a network, and parameter adjustment is also a necessary skill. I would like to share with you the experience and some tips I have accumulated in the project.
PS: Based on personal experience and online experience, everyone makes their own choice.
The following titles will be published in the next few days.
Medical image multi-label classification
How to convert 3D pictures to 2D pictures and labels
Mission introduction
Based on the provided CT thick-slice images, voxel-level 3D labels of the target organs, and classification labels of whether the target organs have traumatic diseases, corresponding algorithms are designed to achieve rapid segmentation of multiple organs in emergency situations and multi-disease screening. The key elements are as follows:
1. Abdominal CT plain thick-slice images, three-dimensional data.
2. Contains voxel-level annotations of the liver, spleen, left kidney, and right kidney and normal/abnormal coding of each organ.
2. First, four binary classifications are needed to determine whether each of the four organs is normal; second, the damaged lesion area of each organ needs to be accurately segmented.
Leveraging 3D Networks
Resnet3D
Leveraging 2D Networks
Resnet2D+thresholding
Join ASLloss
Medical image multi-label segmentation
How to convert 3D pictures to 2D pictures and labels
Leveraging 3D Networks
Refer to UNETR++
https://github.com/ShahinaKK/www.example.com
Leveraging 2D Networks
Deeplabv3plus++, I forgot the reference for this code.
The running results can be very high.
Medical image classification
Mission introduction
Classify CT images as negative and positive for small bowel cancer and draw heatmaps
Overfitting problems encountered and how to solve them
Use a lower layer network
and join dropout