Are you an EPFL student looking for a semester project?

Work with us on data science and visualisation projects, and deploy your project as an app on top of GraphSearch.

Publication

Weakly Supervised Deep Learning Methods for Biomicroscopy

Adrian Shajkofci
EPFL, 2021
EPFL thesis

Abstract

Optical microscopy, an invaluable tool in biology and medicine to observe and quantify cellular function, organ development, or disease mechanisms, requires constant trade-offs between spatial, temporal, and spectral resolution, invasiveness, acquisition time, and post-processing effort. Deep learning technologies have enabled multiple applications that are transforming our day-to-day routines, including the way we approach microscopy. Yet despite the ever-increasing computational power, it is often the lack of labeled training data that is the limiting factor for wide adoption in this domain. Annotating data is often a lengthy and expensive task, since it involves tedious work, generally by skilled experts.

In this thesis, I explored "weakly supervised" learning methods targeted at a variety of applications to enhance microscopy images and extract physical information from a single image. The specificity of these "weakly supervised" methods is the fact that they use very little prior information about the image in order to keep the effort to annotate training data as low as possible. Specifically, I reduced the dimensionality of the learning problem by targeting the experiment towards estimating the parameters of a spatially-variant point-spread function (PSF) model using a convolutional neural network (CNN), which does not require instrument- or object-specific calibration. Using such a model permitted to simulate realistically accurate training data that could be generalized, once the model was trained, to real microscopy images. I extensively benchmarked different network architectures, training datasets and simulation modalities towards the optimal PSF prediction performance and robustness to image degradation.

Starting from the estimated PSF model parameters, I developed a variety of applications, such as a semi-blind spatially-variant deconvolution method for image deblurring and enhancement, a robust and fast microscopy auto-focus, a method for the estimation of the object surface from a single 2D image, and a method for the estimation of the object velocity in a fluid, all of them with minimal need for a priori knowledge about the optical setup.

Official source

https://infoscience.epfl.ch/record/284638?ln=en

About this result

This page is automatically generated and may contain information that is not correct, complete, up-to-date, or relevant to your search query. The same applies to every other page on this website. Please make sure to verify the information with EPFL's official sources.

Related concepts

Related publications

Adrian Shajkofci
EPFL, 2021
EPFL thesis

Abstract

Official source

https://infoscience.epfl.ch/record/284638?ln=en

About this result

Related concepts (20)

Related concepts

Related publications (21)

Related publications

Automatic Segmentation of Nuclei in Histopathology Images using Encoding-Decoding Convolutional Neural Networks

Beril Besbinar, Pascal Frossard, Deniz Sayin Mercadier

Accurate and fast segmentation of nuclei in histopathological images plays a crucial role in cancer research for detection and grading, as well as personal treatment. Despite the important efforts, current algorithms are still suboptimal in terms of speed, adaptivity and generalizability. Popular Deep Convolutional Neural Networks (DCNNs) have recently been utilized for nuclei segmentation, outperforming traditional approaches that exploit color and texture features in combination with shallow classifiers or segmentation algorithms. However, DCNNs need large annotated datasets that require extensive amount of time and expert knowledge. In addition, segmentation results obtained by either traditional or DCNN approaches often require a post-processing step to separate cluttered nuclei. In this paper, we propose a computationally efficient nuclei segmentation framework based on DCNNs exhibiting an encoding-decoding structure. We use a partially-annotated dataset and develop an effective training solution. We also use a weighted background model for network to give more importance to borders of nuclei to overcome the problem of clutters. The abolition of any pre-processing or post-processing step without any compromise on the performance leads to a fast and parameter-free system, which presents important advantages with respect to state-of-the-art.

IEEE2019

Spatially-Variant CNN-Based Point Spread Function Estimation for Blind Deconvolution and Depth Estimation in Optical Microscopy

Michael Stefan Daniel Liebling, Adrian Shajkofci

Optical microscopy is an essential tool in biology and medicine. Imaging thin, yet non-flat objects in a single shot (without relying on more sophisticated sectioning setups) remains challenging as the shallow depth of field that comes with high-resolution microscopes leads to unsharp image regions and makes depth localization and quantitative image interpretation difficult. Here, we present a method that improves the resolution of light microscopy images of such objects by locally estimating image distortion while jointly estimating object distance to the focal plane. Specifically, we estimate the parameters of a spatially-variant Point Spread Function (PSF) model using a Convolutional Neural Network (CNN), which does not require instrument- or object-specific calibration. Our method recovers PSF parameters from the image itself with up to a squared Pearson correlation coefficient of 0.99 in ideal conditions, while remaining robust to object rotation, illumination variations, or photon noise. When the recovered PSFs are used with a spatially-variant and regularized Richardson-Lucy (RL) deconvolution algorithm, we observed up to 2.1 dB better Signal-to-Noise Ratio (SNR) compared to other Blind Deconvolution (BD) techniques. Following microscope-specific calibration, we further demonstrate that the recovered PSF model parameters permit estimating surface depth with a precision of

\mathrm {2 \mu \text {m} }

and over an extended range when using engineered PSFs. Our method opens up multiple possibilities for enhancing images of non-flat objects with minimal need for a priori knowledge about the optical setup.

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC2020

This paper provides a theoretical study of deep neural function approximation in reinforcement learning (RL) with the ϵ-greedy exploration under the online setting. This problem setting is motivated by the successful deep Q-networks (DQN) framework that falls in this regime. In this work, we provide an initial attempt on theoretical understanding deep RL from the perspective of function class and neural networks architectures (e.g., width and depth) beyond the ``linear'' regime. To be specific, we focus on the value based algorithm with the ϵ-greedy exploration via deep (and two-layer) neural networks endowed by Besov (and Barron) function spaces, respectively, which aims at approximating an α-smooth Q-function in a d-dimensional feature space. We prove that, with T episodes, scaling the width

m = \widetilde{\mathcal{O}}(T^{\frac{d}{2\alpha + d}})

and the depth

L=\mathcal{O}(\log T)

of the neural network for deep RL is sufficient for learning with sublinear regret in Besov spaces. Moreover, for a two layer neural network endowed by the Barron space, scaling the width

\Omega(\sqrt{T})

is sufficient. To achieve this, the key issue in our analysis is how to estimate the temporal difference error under deep neural function approximation as the ϵ-greedy exploration is not enough to ensure "optimism". Our analysis reformulates the temporal difference error in an

L^2(\mathrm{d}\mu)

-integrable space over a certain averaged measure μ, and transforms it to a generalization problem under the non-iid setting. This might have its own interest in RL theory for better understanding

\epsilon

-greedy exploration in deep RL.

2022