David J. Ho, PhD
Machine Learning Scientist
Memorial Sloan Kettering Cancer Center
Deep learning-based whole slide image segmentation for efficient and reproducible assistance in pathology
Background: Pathology plays a crucial role to diagnose cancer and to assess its progression from H&E-stained tissue samples. Diagnosis and assessment have been done under microscopes which can be inefficient and subjective. Digitization of glass slides and deep learning-based computational approaches have been investigated to help this process. Especially, semantic segmentation, also known as pixel-wise classification, of whole slide images providing information of location and size of multiple tissue subtypes is a prerequisite step for clinical interpretations. We introduce how tissue segmentation by Deep Multi-Magnification Network (DMMN) can assist cancer diagnosis and assessment in an efficient and reproducible manner.
Methods: DMMN looks at morphological features from multiple magnifications for more accurate segmentation. We trained three DMMN models to help pathologists in clinical settings: (1) a breast model segmenting cancer to screen malignant margin slides where most of margins are generally benign, (2) an osteosarcoma model segmenting viable tumor and necrotic tumor to calculate case-level necrosis ratio from multiple slides for pre-operative treatment response assessment, (3) a lung model segmenting multiple tumor subtypes to find the predominant pattern.
Results: The breast model selected malignant margin slides and highlighted cancer regions with high sensitivity. The osteosarcoma model estimated case-level necrosis ratio with an acceptable error rate comparing to pathologists' manual assessment. The lung model determined the predominant pattern based on multi-class segmentation.
Conclusion: Segmentation models we developed can provide efficient and objective supports to pathologists. We plan to apply these models to clinical settings to reduce pathologists' assessment time and error.
- Understand how tissue segmentation of whole slide images is done by deep learning
- Understand how segmentation can assist diagnosis and assessment of various cancer types
Dr. David Joon Ho is a machine learning scientist at Memorial Sloan Kettering Cancer Center. He received his BS and MS degrees from the University of Illinois at Urbana-Champaign and PhD degree from Purdue University, all in electrical and computer engineering. His research interests include digital and computational pathology, computer vision, and machine learning/deep learning. More specifically, He works on multi-class tissue segmentation of histopathology whole slide images from various cancer types and pursue further analyses such as treatment response assessment, mutation prediction, and treatment response prediction.