The goal of the project is to determine whether the 3rd Molar and the mandibular canal are in contact with each other in CT images.
Data Class: 0(Not exist), 1(No contact), 2(Simple contact), 3(More involve)
Initially, vanila classification, which receives images as input, achieved an accuracy of less than 50%, and the DL classification learning pipeline was subdivided into three stages to increase accuracy;
Learning PIPELINE:
- stage1: the 3rd Molar Binary classification
- 0 or 1,2,3
- Check for the 3rd Molar and create an attention feature map to be used as input in stage3.
- stage2: Canal Line detection & inpainting
- stage3: Multi-output classification
- 1 or 2 or 3ß
[Figure]
stage1:
stage2:
stage1 output: attention feature map
