When a doctor examines a tissue sample under a microscope to diagnose
cancer, the sample must first be chemically stained, which slows the process
and permanently alters the tissue. This thesis addresses a staining-free alternative to that: classifying tissue from infrared spectroscopy (FTIR), where
each image pixel carries a full spectrum instead of a color, from which the
tissue type can be inferred. Such data is large and both spectral and spatial
(a spectral image), which makes deep neural networks, able to exploit both
kinds of information at once, a natural approach. We replicated one of the
first scientific articles to apply convolutional neural networks to this problem; it compares several machine-learning models for the task, and its main
contribution is that spatial information improves predictions, but it releases
neither source code nor all the details needed to reliably reproduce its results.
We recreate the proposed models as faithfully as possible on a proper split
between two physically separate tissue samples, systematically resolve the
missing details, and complement the solution with our own improvements.
Results comparable to those of the original article show that we succeeded
in building equivalent models, including the most complex proposed one, the
spatial-spectral classification model, which reaches a classification accuracy
of 81.56% ± 3.18 (ten repetitions) against the article’s 79.18% ± 1.34.
|
