Dynamically Ranking Urothelial Cells by Malignancy Using Multiple Instance Learning and Attention
Tanay Panja
Lay Summary:
We utilized multiple-instance learning to enhance bladder cancer diagnosis by analyzing and ranking individual cells, even in the early stages, based on their malignancy. This tool allows for more accurate diagnoses and better-targeted treatments, improving care for patients.
Abstract:
Urothelial carcinoma poses significant challenges in diagnosis and treatment due to its heterogeneity, necessitating a dynamic ranking of cancer cells by malignancy for more targeted assessments. In this study, we utilized features from the AutoParis-X project, incorporating cell morphology and deep-learning extracted features with slide diagnoses being labels. Since slide classifications are not indicative of cell malignancy, we utilized multiple-instance learning with bags representing slides and instances being cells. A deep learning-based attention mechanism was employed to assign weights to the feature vector of each cell. Then, we aggregated feature vectors to create an overall slide-level representation. An MLP was then used to predict the malignancy probability of the slide. Finally, cells were ranked based on their relevancy to the malignancy classification using the attention weights. Our results indicate an accuracy of 78.6% and an area under the receiver operating characteristic curve (AUROC) of 0.76, highlighting the potential of our approach to enhance UC diagnostics and treatment strategies.
Q&A:
Bios: Tanay Panja
Program Track: Advanced Research
GitHub Username:
https://github.com/tpanja -Tanay Panja
What was your favorite seminar? Why?
My favorite seminar was about BNEIR. It was interesting to learn about how spatial metal maps are made for specimens, and also to view the rapid progress being made across the nation. -Tanay Panja
If you were to summarize your summer internship experience in one sentence, what would it be?
My summer internship was an invaluable experience that advanced my AI/ML skills and understanding, thanks to exceptional mentorship, especially given the limited resources at Title 1 eligible public school in Michigan. -Tanay Panja