University of Twente Student Theses

Login

Look at this, not that! – Improving the alignment of PIP-Net with domain knowledge

Fobbe, Franziska (2023) Look at this, not that! – Improving the alignment of PIP-Net with domain knowledge.

This is the latest version of this item.

[img] PDF
16MB
Abstract:Interpretable computer vision models like PIPNet can push the application of machine learning models in critical domains like radiology. Based on PIP-Nets prototypes we propose two measures to evaluate a model’s adherence to domain knowledge (which we define as not relying on spurious correlations of medically irrelevant features (shortcuts) and as being focused on regions of interest). We show that the methods are sufficient to discriminate models that adhere more to domain knowledge and can provide an additional dimension to traditional evaluation metrics like accuracy and AUC. By knowing which prototype correspond to shortcuts we can improve the models adherence to domain knowledge by retraining and reinitialising the classification layer of PIP-Net. For regions of interest the same strategy does not work and we propose and discuss a loss term that could improve the results in future research.
Item Type:Essay (Master)
Clients:
Ziekenhuis Group Twente, Hengelo, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:https://purl.utwente.nl/essays/97185
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page