Problems and opportunities in training deep learning software systems

AbstractAccurate and rapid diagnosis of COVID-19 suspected cases plays a crucial role in timely quarantine and medical treatment. Developing a deep learning-based model for automatic COVID-19 detection on chest CT is helpful to counter the outbreak of SARS-CoV-2. A weakly-supervised deep learning-based software system was developed using 3D CT volumes to detect COVID-19. For each patient, the lung region was segmented using a pre-trained UNet; then the segmented 3D lung region was fed into a 3D deep neural network to predict the probability of COVID-19 infectious. 499 CT volumes collected from Dec. 13, 2019, to Jan. 23, 2020, were used for training and 131 CT volumes collected from Jan 24, 2020, to Feb 6, 2020, were used for testing. The deep learning algorithm obtained 0.959 ROC AUC and 0.976 PR AUC. There was an operating point with 0.907 sensitivity and 0.911 specificity in the ROC curve. When using a probability threshold of 0.5 to classify COVID-positive and COVID-negative, the algorithm obtained an accuracy of 0.901, a positive predictive value of 0.840 and a very high negative predictive value of 0.982. The algorithm took only 1.93 seconds to process a single patient’s CT volume using a dedicated GPU. Our weakly-supervised deep learning model can accurately predict the COVID-19 infectious probability in chest CT volumes without the need for annotating the lesions for training. The easily-trained and highperformance deep learning algorithm provides a fast way to identify COVID-19 patients, which is beneficial to control the outbreak of SARS-CoV-2. The developed deep learning software is available at https://github.com/sydney0zq/covid-19-detection.

Download Full-text

A Deep Learning Model for Molecular Label Transfer that Enables Cancer Cell Identification from Histopathology Images

10.1101/2021.03.18.436004 ◽

2021 ◽

Author(s):

Andrew Su ◽

HoJoon Lee ◽

Xiao Tan ◽

Carlos J Suarez ◽

Noemi Andor ◽

...

Keyword(s):

Deep Learning ◽

Initial Step ◽

Classification Systems ◽

Interactive Tools ◽

Histopathologic Evaluation ◽

Histopathological Images ◽

Novel Method ◽

Histopathology Images ◽

Transfer Method ◽

Learning Software

Deep learning cancer classification systems have the potential to improve cancer diagnosis. However, development of these computational approaches depends on prior annotation through a pathologist. This initial step relying on a manual, low-resolution, time-consuming process is highly variable and subject to observer variance. To address this issue, we developed a novel method, H&E Molecular neural network (HEMnet). This two-step process utilises immunohistochemistry as an initial molecular label for cancer cells on a H&E image and then we train a cancer classifier on the overlapping clinical histopathological images. Using this molecular transfer method, we show that HEMnet accurately distinguishes colorectal cancer from normal tissue at high resolution without the need for an initial manual histopathologic evaluation. Our validation study using histopathology images from TCGA samples accurately estimates tumour purity. Overall, our method provides a path towards a fully automated delineation of any type of tumor so long as there is a cancer-oriented molecular stain available for subsequent learning. Software, tutorials and interactive tools are available at: https://github.com/BiomedicalMachineLearning/HEMnet

Download Full-text

Deep Learning in Data-Driven Pavement Image Analysis and Automated Distress Detection: A Review

Data ◽

10.3390/data3030028 ◽

2018 ◽

Vol 3 (3) ◽

pp. 28 ◽

Cited By ~ 23

Author(s):

Kasthurirangan Gopalakrishnan

Keyword(s):

Computer Vision ◽

Image Analysis ◽

Deep Learning ◽

Asset Management ◽

Network Architecture ◽

Crack Detection ◽

Future Research ◽

Deep Convolutional Neural Networks ◽

Pavement Distress ◽

Learning Software

Deep learning, more specifically deep convolutional neural networks, is fast becoming a popular choice for computer vision-based automated pavement distress detection. While pavement image analysis has been extensively researched over the past three decades or so, recent ground-breaking achievements of deep learning algorithms in the areas of machine translation, speech recognition, and computer vision has sparked interest in the application of deep learning to automated detection of distresses in pavement images. This paper provides a narrative review of recently published studies in this field, highlighting the current achievements and challenges. A comparison of the deep learning software frameworks, network architecture, hyper-parameters employed by each study, and crack detection performance is provided, which is expected to provide a good foundation for driving further research on this important topic in the context of smart pavement or asset management systems. The review concludes with potential avenues for future research; especially in the application of deep learning to not only detect, but also characterize the type, extent, and severity of distresses from 2D and 3D pavement images.

Download Full-text

Guest Editor’s Introduction: Software Mathematical Models for Human Understanding

International Journal of Software Engineering and Knowledge Engineering ◽

10.1142/s0218194020020040 ◽

2020 ◽

Vol 30 (10) ◽

pp. 1371-1374

Author(s):

Iaakov Exman

Keyword(s):

Neural Networks ◽

Deep Learning ◽

Systems Design ◽

Difficult Problem ◽

The Other ◽

Software Systems ◽

Special Issue ◽

Software Developers ◽

Human Understanding ◽

Introductory Paper

The unrelenting trend of larger and larger sizes of Software Systems and data has made software comprehensibility an increasingly difficult problem. However, a tacit consensus that human understanding of software is essential for most software related activities, stimulated software developers to embed comprehensibility in their systems’ design. On the other hand, recent empirical successes of Deep Learning neural networks, in several application areas, seem to challenge the tacit consensus: is software comprehensibility a necessity, or just superfluous? This introductory paper, to the 2020 special issue on Theoretical Software Engineering, offers reasons justifying our standpoint on the referred controversy. This paper also points out to specific techniques enabling Human Understanding of software systems relevant to this issue’s papers.

Download Full-text

Human organizational patterns applied to collaborative learning software systems

Computers in Human Behavior ◽

10.1016/j.chb.2014.11.094 ◽

2015 ◽

Vol 51 ◽

pp. 742-751 ◽

Cited By ~ 9

Author(s):

Manuel Kolp ◽

Yves Wautelet

Keyword(s):

Collaborative Learning ◽

Software Systems ◽

Organizational Patterns ◽

Learning Software

Download Full-text

Decision letter: KymoButler, a deep learning software for automated kymograph analysis

10.7554/elife.42288.020 ◽

2018 ◽

Author(s):

Jeff Urbach ◽

Gaudenz Danuser

Keyword(s):

Deep Learning ◽

Learning Software

Download Full-text

Deep learning image recognition enables efficient genome editing in zebrafish by automated injections

10.1101/384735 ◽

2018 ◽

Cited By ~ 2

Author(s):

Maria Lorena Cordero-Maldonado ◽

Simon Perathoner ◽

Kees-Jan van der Kolk ◽

Ralf Boland ◽

Ursula Heins-Marroquin ◽

...

Keyword(s):

Deep Learning ◽

Graphics Processing Unit ◽

Second Step ◽

Processing Unit ◽

Cell Stage ◽

Knock Out ◽

Larval Stages ◽

Learning Software ◽

Graphics Processing ◽

Automated Injection

AbstractOne of the most popular techniques in zebrafish research is microinjection, as it is a rapid and efficient way to genetically manipulate early developing embryos, and to introduce microbes or tracers at larval stages.Here we demonstrate the development of a machine learning software that allows for microinjection at a trained target site in zebrafish eggs at unprecedented speed. The software is based on the open-source deep-learning library Inception v3.In a first step, the software distinguishes wells containing embryos at one-cell stage from wells to be skipped with an accuracy of 93%. A second step was developed to pinpoint the injection site. Deep learning allows to predict this location on average within 42 µm to manually annotated sites. Using a Graphics Processing Unit (GPU), both steps together take less than 100 milliseconds. We first tested our system by injecting a morpholino into the middle of the yolk and found that the automated injection efficiency is as efficient as manual injection (~ 80%). Next, we tested both CRISPR/Cas9 and DNA construct injections into the zygote and obtained a comparable efficiency to that of an experienced experimentalist. Combined with a higher throughput, this results in a higher yield. Hence, the automated injection of CRISPR/Cas9 will allow high-throughput applications to knock out and knock in relevant genes to study their mechanisms or pathways of interest in diverse areas of biomedical research.

Download Full-text