Machine-Guided Biopsy Analysis in Oncology: Facilitating Diagnostic Access and Biomedical Discovery Through Deep Learning
Data collected from cancer patients encompass a wide range of imaging, molecular, and genetic information in an effort to provide more personalized and targeted care. In this the- sis, we outline three case studies examining the role deep learning can play in interpreting and utilizing complex biops...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2024
|
| Online Access: | https://hdl.handle.net/1721.1/155590 https://orcid.org/0000-0002-6529-253X |
| Summary: | Data collected from cancer patients encompass a wide range of imaging, molecular, and genetic information in an effort to provide more personalized and targeted care. In this the- sis, we outline three case studies examining the role deep learning can play in interpreting and utilizing complex biopsy data. In the first case, we employ neural networks to facilitate the detection of high-risk human papillomavirus (HPV) DNA in a point-of-care diagnostic device, facilitating deployment in resource-limited settings. In the second, we develop a comprehensive computational pipeline for cyclic fluorescent microscopy analysis. We iden- tify key immune cell subpopulations in head and neck cancer biopsies to better understand the tumor microenvironment’s influence on disease progression and treatment response. Our third study tackles the analysis of gigapixel-sized digital histology images from breast cancer biopsies. We establish a two-step approach that i) uses self-supervised learning to encode small-scale histological details into robust representations and ii) applies a transformer model to these representations to evaluate larger-scale histological patterns. Our model success- fully distinguished patients with high-risk genetic profiles from histology alone and provided visualization tools to highlight slide regions most closely associated with a high risk of recur- rence. In doing so, we set the stage for deep learning to serve as an alternative to expensive genetic testing as well as a tool for illuminating previously unidentified markers of recurrence risk. These studies underscore deep learning’s versatility in oncology, streamlining the inte- gration of complex datasets into clinical workflows and broadening access to state-of-the-art personalized care. |
|---|