1 Deep Learning Approach for Dynamic Sampling for Multi channel Mass Spectrometry Imaging

2025-04-30 1 0 2.56MB 10 页 10玖币

侵权投诉

Deep Learning Approach for Dynamic Sampling for

Multichannel Mass Spectrometry Imaging

David Helminiak, Member, IEEE, Hang Hu, Julia Laskin, and Dong Hye Ye, Member, IEEE

Abstract—Mass Spectrometry Imaging (MSI), using traditional

rectilinear scanning, takes hours to days for high spatial resolution

acquisitions. Given that most pixels within a sample's field of view

are often neither relevant to underlying biological structures nor

chemically informative, MSI presents as a prime candidate for

integration with sparse and dynamic sampling algorithms. During

a scan, stochastic models determine which locations

probabilistically contain information critical to the generation of

low-error reconstructions. Decreasing the number of required

physical measurements thereby minimizes overall acquisition

times. A Deep Learning Approach for Dynamic Sampling

(DLADS), utilizing a Convolutional Neural Network (CNN) and

encapsulating molecular mass intensity distributions within a

third dimension, demonstrates a simulated 70% throughput

improvement for Nanospray Desorption Electrospray Ionization

(nano-DESI) MSI tissues. Evaluations are conducted between

DLADS and a Supervised Learning Approach for Dynamic

Sampling, with Least-Squares regression (SLADS-LS) and a

Multi-Layer Perceptron (MLP) network (SLADS-Net). When

compared with SLADS-LS, limited to a single m/z channel, as well

as multichannel SLADS-LS and SLADS-Net, DLADS respectively

improves regression performance by 36.7%, 7.0%, and 6.2%,

resulting in gains to reconstruction quality of 6.0%, 2.1%, and

3.4% for acquisition of targeted m/z.

Index Terms— Compressed Sensing, Deep Learning, Machine

Learning, Mass Spectroscopy Imaging, Sparse Sampling

I. INTRODUCTION

ECTILINEAR or raster scanning typically conducts

measurements with statically defined top-down/left-

right movements and remains the most common

imaging pattern for spectroscopy and microscopy,

including Mass Spectrometry Imaging (MSI). MSI measures

molecular distributions at high spatial resolutions and chemical

specificity. However, raster scanning obtains all of the

information within an equipment's Field Of View (FOV),

regardless of that information's relevance to research

objectives. Whether determining the distribution of a targeted

set of molecules or isolating measurements inside the

encapsulated tissue area, avoiding the acquisition of non-

relevant voxels offers potential for significant throughput gains,

even when paired with only basic reconstruction techniques.

Nanospray Desorption Electrospray Ionization (nano-DESI)

[1] serves as an example MSI technology, where such

approaches may be applied, since it can take hours to days for

even a small tissue section to be imaged at high spatial

resolutions. Operationally, nano-DESI MSI moves a sample

This project was funded by award UG3HL145593 from the National Institute

of Health (NIH) Common Fund, through the Office of Strategic Coordination,

as part of the Human BioMolecular Atlas Program's (HuBMAP)

Transformative Technology Development division [22].

Code available at github.com/Yatagarasu50469/SLADS version 0.9.2.

under a dynamic liquid bridge, extracting molecules from its

surface and subsequently analyzing these in a mass

spectrometer. This process yields the spatial distribution of

mass spectra for the sample, quantifying intensities for varying

mass-to-charge (m/z) channels. Contrary to other popular MSI

technologies, particularly Matrix-Assisted Laser Desorption

Ionization (MALDI), nano-DESI requires minimal sample

preparation and can be conducted outside of a vacuum.

However, existing experimental nano-DESI platforms have a

line-bounded geometry constraint, performing independent sets

of measurements along singular rows. While the movement

restriction potentially makes nano-DESI the worst-case MSI

technology for dynamic sparse sampling, the throughput can

still be notably improved through the integration of DLADS.

A. Background

Static scanning patterns using predetermined locations can be

generated for well-defined and consistent structures [2],

alternatively being produced uniformly, randomly, or through

stochastic models [3], [4], [5]. A method growing in popularity,

within Magnetic Resonance Imaging (MRI) and Computed

Tomography (CT), performs random sampling and then relies

on deep learning for in-painting. Although a promising prospect

for improving imaging throughput, there still exist practical and

regulatory requirements for improved result explainability.

An alternative class of algorithms perform dynamic sampling

[6], [7], through the merger of stochastic models and

compressed sensing, where measurement locations are

progressively determined based on data actively being

obtained. A particularly successful implementation is the

Supervised Learning Approach to Dynamic Sampling

(SLADS), first made available in 2017 by Godaliyadda [8].

Thereafter, Scarborough et al. [9] applied SLADS to produce a

20-fold reduction in applied radiation dosage for dynamic X-

Ray crystalline protein acquisition, with only a ~0.1% absolute

difference compared to a full scan. SLADS, or SLADS-LS,

uses a least-squares regression model to produce an Estimated

Reduction in Distortion (ERD), indicating the amount of

entropy that can be removed from a reconstruction for as-of-yet

unmeasured locations, derived from extracted statistical

features. SLADS was generally evaluated by Zhang et al. in

2018, reducing the number of required measurements for

acceptable reconstructions in confocal Raman microscopy [10]

D. Helminiak and D. Ye are with the Department of Electrical and Computer

Engineering, Marquette University, Milwaukee, WI, 53233 USA (e-mail:

david.helminiak@marquette.edu and donghye.ye@marquette.edu).

H. Hu and J. Laskin are with the Chemistry Department, Purdue University,

West Lafayette, IN, 47907 USA (e-mail: hu518@purdue.edu and

jlaskin@purdue.edu).

by 6-fold, ~60-95% for Electron Back Scatter Diffraction

(EBSD) [11], up to 90% for Energy Dispersive X-Ray

Spectroscopy (EDS) [12], and between ~60-90% for metal

dendrite sampling when combined with Hierarchical Gaussian

Mixture Models (HGMMs) in a model named Unsupervised-

SLADS (U-SLADS) [13]. During this time, Zhang et al. also

published a multi-model study [14], examining the replacement

of the least-squares regression with either a Support Vector

Machine (SVM), or a Multi-Layer Perceptron (MLP) network.

The MLP, identified as SLADS-Net, improved generalization

between training and testing with dissimilar data. An alternate

approach was published in 2020 by Grosche et al. [15], using a

Probabilistic Approach to Dynamic Image Sampling (PADIS)

for single-channel Scanning Electron Microscope (SEM)

images, relying on a probability mass function for selection of

scanning locations. PADIS outperformed both SLADS-LS and

SLADS-Net, which both tended to oversample structural edges.

SLADS has been limited to processing 2-Dimensional (2D)

images, where a third dimension (3D) risks obfuscation of

channel-specific details [16]. Further, SLADS implementations

have focused on structural information, where even if the data

was continuous, non-structural areas were homogeneous. MSI

data is highly heterogeneous, necessitating more complex

models to effectively leverage captured information. An initial

study was conducted by Helminiak et al. in 2021 [17], [18] on

the feasibility of applying SLADS-LS, SLADS-Net, and a new

Deep Learning Approach for Dynamic Sampling (DLADS),

using a sequential Convolutional Neural Network (CNN), with

nano-DESI MSI tissues. The work reduced the data into 2D by

averaging 10 arbitrary mass ranges, empirically observed to

contain desirable information. Therein, DLADS reduced the

number of required measurements to achieve an acceptable

reconstruction for the averaged m/z by 70-80% and

demonstrated a 14-46% performance improvement over single-

channel SLADS-LS and SLADS-Net. However, where the

2021 study used samples that had already been post-processed.

Combination of DLADS with nano-DESI MSI can only be

realized after these steps have been re-designed and integrated

for actual scans. First, there exists a line-bounded movement

constraint on the acquisition probe, where measurements can

only be performed along a single indicated line in each scanning

cycle. Next, the dimensionality of the measurable locations

along that line depend on the actual scan and acquisition rates,

which can be variable when using equipment automatic gain

control. Additional inconsistencies include the physical start

and stop locations for each row, as well as the specific m/z

where intensities are measured.

Fig. 1. Overview of dynamic sampling models' training and testing procedures. During training, fully measured tissue sample data are sparsely

sampled and reconstructed, from which corresponding ground

truth RD are determined. This data trains machine learning models to produce an

ERD output, given currently known information, thereby guiding progressive measureme

nts during simulated and experimental acquisitions.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

1DeepLearningApproachforDynamicSamplingforMultichannelMassSpectrometryImagingDavidHelminiak,Member,IEEE,HangHu,JuliaLaskin,andDongHyeYe,Member,IEEEAbstract—MassSpectrometryImaging(MSI),usingtraditionalrectilinearscanning,takeshourstodaysforhighspatialresolutionacquisitions.Giventhatmostpixelswithina...

展开>> 收起<<

1 Deep Learning Approach for Dynamic Sampling for Multi channel Mass Spectrometry Imaging.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

1 Deep Learning Approach for Dynamic Sampling for Multi channel Mass Spectrometry Imaging

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: