Appendix 1: Content Summaries of Selected Best Papers for the 2021 IMIA Yearbook,
Section Sensors, Signals, and Imaging Informatics (CB)
Gemein LAW, Schirrmeister RT, Chrabąszcz P, Wilson D, Boedecker J, Schulze-Bonhage
A, Hutter F, Ball T
Machine-learning-based diagnostics of EEG pathology
Neuroimage 2020 Oct 15;220:117021
The analysis of clinical electroencephalograms (EEGs) is a time-consuming and demanding
process and requires years of training. The development of algorithms for automatic
EEG diagnosis, such as machine learning (ML) methods, could be a tremendous benefit
to clinicians in analyzing EEGs. In this work, end-to-end decoding using deep neural
networks was compared with feature-based decoding using a large set of features. Approximately
3,000 recordings from the Temple University Hospital EEG Corpus (TUEG) study were
used, representing the largest publicly available collection of EEG recordings to
date. For feature-based pathology decoding, Random Forest (RF), Support Vector Machine
(SVM), Riemannian geometry (RG), and Auto-Skill Classifier (ASC) were used, while
three types of convolutional neural networks (CNN) were applied for end-to-end pathology
decoding: the 4-layer ConvNet architecture Braindecode Deep4 ConvNet (BD-Deep4), Braincode
(BD) and TCN. The main result of this study was that the EEG pathology decoding accuracy
is in a narrow range of 81-86%, also compared to a wide range of analysis strategies,
network archetypes, network architects, feature-based classifiers and ensembles, and
datasets. Based on the feature visualizations, features extracted in the theta and
delta regions of temporal electrode positions were considered informative. Feature
correlation analysis showed strong correlations of features extracted at different
electrode positions. Besides the fact that there is no statistical evidence that the
deep neural networks studied perform better than the feature-based approach, this
work presents that a somewhat elaborate feature-based approach can be used to achieve
similar decoding results as deep end-to-end methods. The authors recommend decoding
specific labels to avoid the consequences of label noise in decoding EEG pathology.
This work provides a remarkable and objective comparison between deep learning and
feature-based methods based on numerous experiments, including cross-validation, bootstrapping,
and input signal perturbation strategies.
Karimi D, Dou H, Warfield SK, Gholipour A
Deep learning with noisy labels: Exploring techniques and remedies in medical image
analysis
Med Image Anal 2020 Oct;65:101759
Label noise is unavoidable in many medical image datasets. It can be caused by limited
attention or expertise of the human annotator, the subjective nature of labeling,
or errors in computerized labeling systems. This is especially concerning for medical
applications where datasets are typically small, labeling requires domain expertise
and suffers from high inter- and intra-observer variability, and erroneous predictions
may influence decisions directly impacting human health. The authors reviewed the
state-of-the-art label noise handling in deep learning and investigated how these
methods were applied to medical image analysis. Their key recommendations to account
for label noise are: label cleaning and pre-processing, adaptions on network architectures,
the use of label-noise-robust loss functions, re-weighting data, label consistency
checks, and the choice of training procedures. They underpin their findings with experiments
on three medical datasets where label noise was introduced by the systematic error
of a human annotator, the inter-observer variability, or the noise generated from
an algorithm. Their results suggest a careful curation of data for training deep learning
algorithms for medical image analysis. Furthermore, the authors recommend integrating
label noise analyses in development processes for robust deep learning models.
Langner T, Strand R, Ahlström H, Kullberg J
Large-scale biometry with interpretable neural network regression on UK Biobank body
MRI
Sci Rep 2020 Oct 20;10(1):17752
This work presents a novel neural network approach for image-based regression to infer
64 biological metrics (beyond age) from neck-to-knee body MRIs with relevance for
cardiovascular and metabolic diseases. Image data were collected from the UK Biobank
study, linked to extensive metadata comprising non-imaging properties such as measurements
of body composition by dual-energy X-ray absorptiometry (DXA) imaging, patient-related
parameters, i.e., age, sex, height and weight, and additional biomarkers for cardiac
health including pulse rate, accumulated fat in the liver and grip strength. The authors
adapted and optimized a previously presented regression pipeline for age estimation
using a ResNet50 architecture, not requiring any manual intervention or direct access
to reference segmentations. Based on 31,172 magnetic resonance imaging (MRI) scans,
the neural network was trained and cross-validated on simplified, two-dimensional
representations of the MR images and evaluated by generated predictions and saliency
maps for all examined properties. The work is noteworthy for its extensive validation
of both the whole framework and predictions, demonstrating a robust performance and
outperforming linear regression baseline in all applied cases. Saliency analysis showed
that the developed neural network accurately targets specific body regions, organs,
and limbs of interest. The network can emulate different modalities, including DXA
or atlas-based MRI segmentation, and on average, correctly targets specific structures
on either side of the body. The authors impressively demonstrated how convolutional
neural network regression could effectively be applied in MRI and offer a first valuable,
fully automated approach to measure a wide range of important biological metrics from
single neck-to-knee body MRIs.
Saito H, Aoki T, Aoyama K, Kato Y, Tsuboi A, Yamada A, Fujishiro M, Oka S, Ishihara
S, Matsuda T, Nakahori M, Tanaka S, Koike K, Tada T
Automatic detection and classification of protruding lesions in wireless capsule endoscopy
images based on a deep convolutional neural network
Gastrointest Endosc 2020 Jul;92(1):144-151.e1
Wireless capsule endoscopy (WCE) is an established examination method for the diagnosis
of small-bowel diseases. Automated detection and classification of protruding lesions
of various types from WCE images is still challenging because it takes 1 to 2 hours
on average for a correct diagnosis by a physician. In this work, a deep neural network
architecture, termed single shot multibox detector (SSD) based on a deep convolutional
neural network (CNN) structure with 16 or more layers, was trained on 30,584 WCE images
from 292 patients collected from multiple centers and tested on an independent set
of 17,507 images from 93 patients, including 7507 images of protruding lesions from
73 patients. All regions showing protruding lesions were manually annotated by six
independent expert endoscopists, representing the ground truth for training the network.
The CNN performance was evaluated by a ROC analysis, revealing an AUC of 0.911, a
sensitivity of 90.7%, and a specificity of 79.8% at the optimal cut-off value of 0.317
for the probability score. In a subanalysis of the categories of protruding lesions,
the sensitivities appeared between 77.0% and 95.8% for the detection of polyps, nodules,
epithelial tumors, submucosal tumors, and venous structures, respectively. In individual
patient analyses, the detection rate of protruding lesions was 98.6%. The rates of
concordance of the labeling by the CNN and three expert endoscopists were between
42% and 83% for the different morphological structures. A false positive/negative
error analysis was reported, indicating some limitations of the current approach in
terms of an imbalanced number of cases, color diversity, and variation of structures
in the images. The work is notable for its excellent clinical applicability using
a new computer-aided system with good diagnostic performance to detect protruding
lesions in small-bowel capsule endoscopy.
