WPI - Computer Science Department , MS Thesis Presentation, Grant Perkins " Pressure Injury Stage Image Classification using Cross-Shaped Window Attention Vision Transformers"
11:00 am to 12:00 pm
WPI – Computer Science Department
Friday, December 1, 2023
Time: 11:00 a.m. – 12:00 p.m.
Location: Fuller Labs 311
Advisor: Prof. Emmanuel Agu
Reader: Prof. Fabricio Murai
Pressure injuries affect up to three million people globally, leading to extended hospital stays and financial hardship for patients and their families. Proper treatment for a pressure injury depends on its stage (severity). Unfortunately, nurses have been found to be less than 70% accurate when diagnosing the stage of a pressure injury, resulting in improper treatment, and delaying the healing process for those misdiagnosed.
To improve this situation, researchers have utilized deep learning models to classify the stage of a pressure injury from an image. Deep learning models struggle to learn from pressure injury image datasets for two reasons: limited samples and high intra-class variation with low inter-class variation. To address these challenges, we propose adapting the Cross-Shaped Window (CSWin) transformer to the task of pressure injury classification. Specifically, cross-shaped window self-attention decreases intra-class variation and increase inter-class variation more efficiently than previous approaches.
We also utilize a model fine-tuning methodology to improve model robustness by pre-training the CSWin transformer on a larger and generalized dataset before fine-tuning it on our smaller pressure injury dataset, facilitating effective pressure injury stage classification from few samples. In rigorous evaluation, CSWin achieved an accuracy of 78.5%, outperforming the accuracy of prior state-of-the-art approaches by more than 8%. Deployment of the CSWin transformer in a hospital setting could improve the diagnostic accuracy by non-expert medical staff, directly improving the health of patients suffering from pressure injuries.