Self-knowledge distillation for first trimester ultrasound saliency prediction
Self-knowledge distillation (SKD) is a recent and promising machine learning approach where a shallow student network is trained to distill its own knowledge. By contrast, in traditional knowledge distillation a student model distills its knowledge from a large teacher network model, which involves...
| Κύριοι συγγραφείς: | , , , , |
|---|---|
| Μορφή: | Conference item |
| Γλώσσα: | English |
| Έκδοση: |
Springer
2022
|
| Περίληψη: | Self-knowledge distillation (SKD) is a recent and promising machine learning approach where a shallow student network is trained to distill its own knowledge. By contrast, in traditional knowledge distillation a student model distills its knowledge from a large teacher network model, which involves vast computational complexity and a large storage size. Consequently, SKD is a useful approach to model medical imaging problems with scarce data. We propose an original SKD framework to predict where a sonographer should look next using a multi-modal ultrasound and gaze dataset. We design a novel Wide Feature Distillation module, which is applied to intermediate feature maps in the form of transformations. The module applies a more refined feature map filtering which is important when predicting gaze for the fetal anatomy variable in size. Our architecture design includes ReSL loss that enables a student network to learn useful information whilst discarding the rest. The proposed network is validated on a large multi-modal ultrasound dataset, which is acquired during routine first trimester fetal ultrasound scanning. Experimental results show the novel SKD approach outperforms alternative state-of-the-art architectures on all saliency metrics. |
|---|