Security and Privacy Team Seminar (Talk by Dr. Joshua Butke , University of Tsukuba, Department of Computer Science, Machine Learning and Data Mining Lab).

This is an online seminar.
Registration is required.

【Speaker】Dr.Joshua Butke
University of Tsukuba, Department of Computer Science, Machine Learning and Data Mining Lab

Title：Paradigm Shifts in (label-free) Computational Pathology

Abstract：Computational pathology is mainly concerned with the medical image
analysis of patient specimens for the study of disease. The development
of novel decision support systems is one way to further personalized
medicine efforts and enable differential diagnosis systems for early
detection and characterization of diseases such as colorectal or urothelial
carcinoma. With recent advances in (label-free) digital pathology,
conventional glass slides carrying patient specimens can now be digitized
into gigapixel whole slide images.
Analyzing those images is mostly facilitated by deep learning-based
artificial intelligence approaches. Here, a key challenge is the lack of
large-scale annotations which serve as the ground truth to learn from.
While many methods were based on strongly supervised learning with
relatively small amounts of intricately labeled data, there is now a shift
towards much larger datasets with weak labeling due to cheaper storage
and computational resources, but also because of the large efforts
required to obtain professional annotations.
My thesis presents novel deep learning approaches from the realm of
representation and multiple instance learning to characterize cancer in
cells and tissues from a spectral and morphological perspective. Infrared
microscopy provides unstained, spatially-resolved spectral images, with
biochemical fingerprints at each pixel in an image, enabling the discovery
of underlying changes in genome, proteome or metabolome. But also light
microscopic morphological information is analyzed, comparable to
standard clinical pathology practice, considering form and structure of
cells and tissues via staining dyes.

By either learning robust representations of high-level features or finding
key elements within the data instances automatically and without
additional annotation efforts, whole slide images can be analyzed
effectively. This domain-adaptable characterization can aid clinical
diagnosis and therapy recommendations by providing a new, non-
invasive alternative of urine cytology based on morphological changes
that differentiate cancerous urothelial cells from non-cancerous ones.
Similarly, spectral changes in colorectal tissue can reveal the
microsatellite status, an important indicator that recommends patients for
specialized immunotherapy, without needing any laborious and not widely
employed staining procedures.