Deep representation learning for domain adaptable classification of infrared spectral imaging data.

MOTIVATION: Applying infrared microscopy in the context of tissue diagnostics heavily relies on computationally preprocessing the infrared pixel spectra that constitute an infrared microscopic image. Existing approaches involve physical models, which are non-linear in nature and lead to classifiers that do not generalize well, e.g. across different types of tissue preparation. Furthermore, existing preprocessing approaches involve iterative procedures that are computationally demanding, so that computation time required for preprocessing does not keep pace with recent progress in infrared microscopes which can capture whole-slide images within minutes. RESULTS: We investigate the application of stacked contractive autoencoders as an unsupervised approach to preprocess infrared microscopic pixel spectra, followed by supervised fine-tuning to obtain neural networks that can reliably resolve tissue structure. To validate the robustness of the resulting classifier, we demonstrate that a network trained on embedded tissue can be transferred to classify fresh frozen tissue. The features obtained from unsupervised pretraining thus generalize across the large spectral differences between embedded and fresh frozen tissue, where under previous approaches separate classifiers had to be trained from scratch. AVAILABILITY AND IMPLEMENTATION: Our implementation can be downloaded from https://github.com/arnrau/SCAE_IR_Spectral_Imaging. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

A framework for falsifiable explanations of machine learning models with an application in computational pathology.

In recent years, deep learning has been the key driver of breakthrough developments in computational pathology and other image based approaches that support medical diagnosis and treatment. The underlying neural networks as inherent black boxes lack transparency and are often accompanied by approaches to explain their output. However, formally defining explainability has been a notorious unsolved riddle. Here, we introduce a hypothesis-based framework for falsifiable explanations of machine learning models. A falsifiable explanation is a hypothesis that connects an intermediate space induced by the model with the sample from which the data originate. We instantiate this framework in a computational pathology setting using hyperspectral infrared microscopy. The intermediate space is an activation map, which is trained with an inductive bias to localize tumor. An explanation is constituted by hypothesizing that activation corresponds to tumor and associated structures, which we validate by histological staining as an independent secondary experiment.

A representation learning approach for recovering scatter-corrected spectra from Fourier-transform infrared spectra of tissue samples.

Infrared spectra obtained from cell or tissue specimen have commonly been observed to involve a significant degree of scattering effects, often Mie scattering, which probably overshadows biochemically relevant spectral information by a nonlinear, nonadditive spectral component in Fourier transform infrared (FTIR) spectroscopic measurements. Correspondingly, many successful machine learning approaches for FTIR spectra have relied on preprocessing procedures that computationally remove the scattering components from an infrared spectrum. We propose an approach to approximate this complex preprocessing function using deep neural networks. As we demonstrate, the resulting model is not just several orders of magnitudes faster, which is important for real-time clinical applications, but also generalizes strongly across different tissue types. Using Bayesian machine learning approaches, our approach unveils model uncertainty that coincides with a band shift in the amide I region that occurs when scattering is removed computationally based on an established physical model. Furthermore, our proposed method overcomes the trade-off between computation time and the corrected spectrum being biased towards an artificial reference spectrum.