-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathview_classification.html
More file actions
581 lines (526 loc) · 44.7 KB
/
Copy pathview_classification.html
File metadata and controls
581 lines (526 loc) · 44.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
<!DOCTYPE html>
<html lang="en">
<!-- ======= Head Section - DON'T CHANGE ======= -->
<head>
<meta charset="utf-8">
<meta content="width=device-width, initial-scale=1.0" name="viewport">
<title>View Classification</title>
<link href="assets/img/logo.png" rel="icon">
<!-- Google Fonts -->
<link href="https://fonts.googleapis.com/css?family=Open+Sans:300,300i,400,400i,600,600i,700,700i|Raleway:300,300i,400,400i,500,500i,600,600i,700,700i|Poppins:300,300i,400,400i,500,500i,600,600i,700,700i" rel="stylesheet">
<!-- Vendor CSS Files -->
<link href="assets/vendor/aos/aos.css" rel="stylesheet">
<link href="assets/vendor/bootstrap/css/bootstrap.min.css" rel="stylesheet">
<link href="assets/vendor/bootstrap-icons/bootstrap-icons.css" rel="stylesheet">
<link href="assets/vendor/boxicons/css/boxicons.min.css" rel="stylesheet">
<link href="assets/vendor/glightbox/css/glightbox.min.css" rel="stylesheet">
<link href="assets/vendor/remixicon/remixicon.css" rel="stylesheet">
<link href="assets/vendor/swiper/swiper-bundle.min.css" rel="stylesheet">
<!-- Main CSS File -->
<link href="assets/css/style.css" rel="stylesheet">
</head>
<!-- ======= Body Section ======= -->
<body>
<!-- ======= Navigation Bar - DON'T CHANGE ======= -->
<header id="header" class="fixed-top">
<div class="container d-flex align-items-center justify-content-between">
<a href="index.html" class="logo"><img src="assets/img/logo.png" alt="" class="img-fluid"></a>
<nav id="navbar" class="navbar">
<ul>
<li><a class="nav-link scrollto" href="index.html">Home</a></li>
<li><a class="nav-link scrollto" href="index.html#projects">Projects</a></li>
<li><a class="nav-link scrollto" href="index.html#team">People</a></li>
<li><a class="nav-link scrollto" href="index.html#resources">Resources</a></li>
<li><a class="nav-link scrollto" href="index.html#funding">Funding</a></li>
<li><a class="nav-link scrollto" href="index.html#partners">Partners</a></li>
<li><a class="nav-link scrollto" href="index.html#workwithus">Vacancies</a></li>
<li><a class="nav-link scrollto" href="index.html#contact">Contact</a></li>
</ul>
<i class="bi bi-list mobile-nav-toggle"></i>
</nav>
</div>
</header><!-- End Navigation Bar -->
<!-- ======= Title Section ======= -->
<section id="project-title" class="d-flex align-items-center">
<div class="container position-relative" data-aos="fade-up" data-aos-delay="100">
<div class="row">
<div class="col-12 text-left">
<h1>Neural Architecture Search of Echocardiography View Classifiers</h1>
<!-- <h2><i>Neda Azarmehr <sup>1,2</sup>, Xujiong Ye <sup>2</sup>, James P Howard <sup>3</sup>, Elisabeth S Lane <sup>1</sup>, Robert Labs <sup>1</sup>, Matthew J Shun-shin <sup>3</sup>, Graham D Cole <sup>3</sup>, Luc Bidaut <sup>2</sup>, Darrel P Francis <sup>3</sup>, Massoud Zolgharni <sup>1,3</sup><i></h2>
<h3><sup>1</sup> School of Computing and Engineering, University of West London, London, United Kingdom<h3>
<h3><sup>2</sup> School of Computer Science, University of Lincoln, Lincoln, United Kingdom<h3>
<h3><sup>3</sup> National Heart and Lung Institute, Imperial College, London, United Kingdom<h3> -->
</div>
</div>
</div>
</section><!-- End Title -->
<!-- Main Webpage Section -->
<main id="project-main">
<!-- ======= Project Body Section ======= -->
<section id="project-body">
<!-- ======= Project Introduction ======= -->
<div class="project-intro">
<div class="container">
<div class="row justify-content-left">
<div class="col-12">
<div align="justify">
<hr>
<h2><b>Project Introduction</h2>
<h2>Purpose:</h2>
<h3>Echocardiography is the most commonly used modality for assessing the heart in clinical practice. In an echocardiographic exam, an ultrasound probe samples the heart from different orientations and positions, thereby creating different viewpoints for assessing the cardiac function. The determination of the probe viewpoint forms an essential step in automatic echocardiographic image analysis.</h3>
<h2>Approach:</h2>
<h3>In this study, convolutional neural networks are used for the automated identification of 14 different anatomical echocardiographic views (larger than any previous study) in a dataset of 8,732 videos acquired from 374 patients. Differentiable architecture search approach was utilised to design small neural network architectures for rapid inference while maintaining high accuracy. The impact of the image quality and resolution, size of the training dataset, and number of echocardiographic view classes on the efficacy of the models were also investigated.</h3>
<h2>Results:</h2>
<h3>In contrast to the deeper classification architectures, the proposed models had significantly lower number of trainable parameters (up to 99.9\% reduction), achieved comparable classification performance (accuracy 88.4-96.0\%, precision 87.8-95.2\%, recall 87.1-95.1\%) and real-time performance with inference time per image of 3.6-12.6ms.</h3>
<h2>Conclusion:</h2>
<h3>Compared with the standard classification neural network architectures, the proposed models are faster and achieve comparable classification performance. They also require less training data. Such models can be used for real-time detection of the standard views.</h3>
<br>
</div>
</div>
</div>
<!-- <div class="text-left">
<a class="btn btn-primary" href="https://doi.org/10.1117/1.JMI.8.3.034002" role="button">Downl the paper</a>
</div> -->
</div>
</div>
<!-- ======= Project Dataset ======= -->
<div class = "project-dataset">
<div class="container project-container">
<div class="row justify-content-left">
<div class="col-12 col-md-12 col-lg-6 col-xl-6">
<img class="img-fluid" src="assets/img/projects/view_classification/dataset1.png" alt="dataset">
<h3> Fig. 1 Distribution of data in the training, validation and test dataset; values show the number of
frames in a given class </h3>
</div>
<div class="col-12 col-md-12 col-lg-6 col-xl-6">
<div align="justify">
<h2><b>Dataset</b></h2>
<h3>A random sample of 374 echocardiographic examinations of different patients and performed between 2010 and 2020 was extracted from Imperial College Healthcare NHS Trust’s echocardiogram database. The acquisition of the images was performed by experienced echocardiographers and according to standard protocols, using ultrasound equipment from GE and Philips manufacturers.</h3>
<h3>Ethical approval was obtained from the Health Regulatory Agency (Integrated Research Application System identifier 243023). Only studies with full patient demographic data and without intravenous contrast administration were included. Automated anonymization was performed to remove all patient-identifiable information.</h3>
<h3>The videos were annotated manually by an expert cardiologist (JPH), categorising each video into one of 14 classes. Videos thought to show no identifiable echocardiographic features, or which depicted more than one view, were excluded. Altogether, this resulted in 9,098 echocardiographic videos. Of these, 8,732 (96.0\%) videos could be classified as one of the 14 views by the human expert. The remaining 366 videos were not classifiable as a single view, either because the view changed during the video loop, or because the images were completely unrecognisable. The cardiologist's annotations of the videos were used as the ground truth for all constituent frames of that video.</h3>
<h3>DICOM-formatted videos of varying image sizes (480*640, 600*800, and 768*1024 pixels)} were then split into constituent frames, and three frames were randomly selected from each video to represent arbitrary stages of the heart cycle, resulting in 41,321 images. The dataset was then randomly split into training (24791 images), validation (8265 images), and testing (8265 images) sub-datasets in a 60:20:20 ratio. Each sub-datasets contained frames from separate echo studies to maintain sample independence.</h3>
<h3>The relative distribution of echo view classes labelled by the expert cardiologist is displayed in figure and indicates an imbalanced dataset, with a ratio of 3\% (Subcostal-IVC view as the least represented class) to 13\% (PSAX-AV view as the dominant view).</h3>
<br>
</div>
<div class="text-left">
<!-- <a class="btn btn-primary" href="#contact" role="button">Download the dataset</a>-->
</div>
</div>
</div>
</div>
<!-- ======= Project Architecture ======= -->
<div class = "project-architecture">
<div class="container">
<hr>
<div class="row justify">
<div class="col-12">
<h2><b>Network Architecture</b></h2>
</div>
<div class="col-12 col-md-12 col-lg-6 col-xl-6">
<div align="justify">
<h3><b> DARTS Method:</b> <P>DARTS method consists of two stages: architecture search and architecture evaluation. <p>Given
the input images, it first embarks on an architecture search to explore for a computation cell (a small unit of convolutional layers) as the building block of the neural network architecture.
<p class="text-justify"> Figure 2 provides an overview of the DARTS network architecture. <p> After the architecture search phase is complete and the optimal cell is obtained based on its
validation performance, the final architecture could be formed from one cell or a sequential stack
of cells. The weights of the optimal cell learned during the search stage are then discarded, and
are initialized randomly for retraining the generated neural network model from scratch. <p> A cell, shown in figure 2, is an ordered sequence of several nodes in which one or multiple
edges meet. Each node C(i) represents a feature map in convolutional networks. Each edge (i,j) is associated with some operation O(i,j), transforming the node C(i) to C(j)
. This could be a combination of several operations, such as convolution, max-pooling, and ReLU. Each intermediate node C(i) is computed based on all of its predecessors. Instead of applying a single operation (e.g., 5 × 5 convolution), and evaluating all possible
operations independently (each trained from scratch), DARTS places all candidate operations on each edge (e.g., 5 × 5 convolution, 3 × 3 convolution, and max-pooling represented in figure to the right.
red, blue, and green lines, respectively). This allows sharing and training their weights in a single
process. The task of learning the optimal cell is effectively finding the optimal placement of
operations at the edges.
The actual operation at each edge is then a linear combination of all candidate operations
O(i,j), weighted by the softmax output of the architecture parameters α(i,j).
<p>
</h3>
</div>
</div>
<div class="col-12 col-md-12 col-lg-6 col-xl-6 text-center ">
<br>
<img class="img-fluid" src="assets\img\projects\view_classification\DARTA.PNG" alt="DARTS" width="450px" height="500px">
<div align="center">
<h3> Fig. 2 Schematic of a DARTS cell. </h3>
</div>
</div>
</div>
</div>
</div>
<!-- ======= Project Implementation ======= -->
<div class = "project-implementation">
<div class="container">
<div class="row justify-content-left">
<div class="col-12">
<div align="justify">
<hr>
<h2><b>Implementation:</b> </h2>
<h3><p class="text-justify">PyTorch was used to implement the models. For the computationally intensive stage of
architecture search, a GPU server equipped with 4 NVIDIA TITAN RTX GPUs with 64 GB
of memory was rented. <p>For the subsequent training of the searched networks and also the standard models, the utilized GPU was an Nvidia QUADRO M5000 with 8 GB of memory, representing a more widely accessible hardware for real-time applications. Inference time (latency
time for classifying each image) was also estimated with the trained models running on the GPU.
To this end, a total of 100 images were processed in a loop, and the average time was recorded.
<p>All training/prediction computations were carried using identical hardware and software resources, allowing for a fair comparison of computational time-efficiency between all network models
investigated in this study.
The number of trainable parameters in the model, as well as the training time per epoch was
also recorded for all CNN networks.
</h3>
<div class="text-left">
<br>
<!-- <a class="btn btn-primary" href="" role="button">Download the code</a> -->
</div>
</div>
</div>
</div>
</div>
</div>
<div class = "project-Training Parameters">
<div class="container">
<div class="row justify-content-left">
<div class="col-12">
<div align="justify">
<hr>
<h2><b>Models Training Parameters</b></h2>
<h3><p class="text-justify">Training occurred subsequently, using annotations provided by the expert cardiologist. It was carried out independently for each of the four different image sizes of 32 × 32, 64 × 64, 96 × 96,
and 128 × 128 pixels. Identical training, validation, and testing datasets were used in all network
models. The validation dataset was used for early stopping to avoid redundant training and overfitting. Each model was trained until the validation loss plateaued. The test dataset was used for
the performance assessment of the final trained models. The DARTS models were kept blind to
the test dataset during the stage of architecture search.
<p>Adam optimizer with a learning rate of 10−4 and a maximum number of 800 epochs was used
for training the models. The cross-entropy loss was used as the networks objective function. For
training the DARTS model, a learning rate of 0.1 deemed to be a better compromise between
speed of learning and precision of result and was therefore used. A batch size of 64 or the maximum which could be fitted on the GPU (if <64) was employed.
<p>The dataset is fairly imbalanced with unequal distribution of
different echo views. To prevent potential biases towards more dominant classes, we used online
batch selection where the equal number of samples from each view were randomly drawn (by
over-sampling of underrepresented classes). This led to training on a balanced dataset representing all classes in every epoch. An epoch was still defined as the number of iterations required for
the network to meet all images in the training dataset.</h3>
<div class="text-left">
<br>
<!-- <a class="btn btn-primary" href="" role="button">Download the code</a> -->
</div>
</div>
</div>
</div>
</div>
</div>
<!-- ======= Project Evaluation ======= -->
<div class = "project-evaluation">
<div class="container">
<div class="row justify-content-left">
<div class="col-12">
<div align="justify">
<hr>
<div class="text-justify">
<h2><b>Evaluation metrics</b></h2>
<h3>Several metrics were employed to evaluate the performance of the investigated models in this
study. Overall accuracy was calculated as the number of correctly classified images as a fraction
of the total number of images. <p>Macro average precision and recall (average overall views of perview measures) were also computed.<p>F1 score was calculated as the harmonic mean of the precision and recall. Since this study is a multi-class problem, F1 score was the weighted average,
where the weight of each class was the number of samples from that class..</h3>
</div>
</div>
</div>
</div>
</div>
</div>
<!-- ======= Project Results ======= -->
<div class = "project-results">
<div class="container align-middle">
<hr>
<div class="row">
<div class="col-12 col-lg-6 col-md-6 text-justify">
<img class="img-fluid" src="assets\img\projects\view_classification\Arch2.PNG" alt="Arch search">
<div class="text-justify">
<div align="justify">
<h3> Fig. 3 Optimal normal and reduction cells for the input image size of 128 × 128 pixels</h3>
<h2><p> <br> <b>2. View Classification:</b></h2>
<h3> Results for five different network topologies and different image sizes are provided in Table 1.
Despite having significantly fewer trainable parameters, the two DARTS models showed competitive results when compared with the standard classification architectures (i.e., VGG16,
ResNet18, and DenseNet201). <br><p><br>The 2-cell-DARTS model, with only ∼0.5-m trainable parameters, achieves the best accuracy (93% to 96%), precision (92.5% to 95.2%), and recall (92.3% to
95.1%) among all networks and across all input image resolutions. Deeper standard neural networks, if employed for echo view detection, would therefore be significantly redundant, with up
to 99% redundancy in trainable parameters.<br>
<p>On the other hand, while maintaining a comparable accuracy to standard network topologies,
the 1-cell-DARTS model has ≤0.09 m trainable parameters and the lowest inference time
amongst all models and across different image resolutions (range 3.6 to 7.2 ms). <p>This would
allow processing about 140 to 280 frames per second, thus making real-time echo view classification feasible.
Compared with manual decision making, this is a significant speedup. Although the identification of the echo view by human operators is almost instantaneous (at least for easy cases),
the average time for the overall process of displaying/identifying/recording the echo view takes
several seconds.
<p>Having fewer trainable parameters, both DARTS models also exhibit faster convergence and
shorter training time per epoch than standard deeper network architectures: 157 116 s versus
622 576 s, respectively, for the training dataset we used.</h3>
<p><br><br> <br><h3> The confusion matrix for the 2-cell-DARTS model and image resolution of 128 × 128 pixels
is provided in Fig. 4. The errors appear predominantly clustered between a certain pair of views
which represent anatomically adjacent imaging planes. The A5CH view proves to be the hardest
one to detect (accuracy of about 80%), as the network is confused between this view and other
apical windows. <p>This is in line with previous observations that the greatest challenge lies in
distinguishing between the various apical views.
<p>Interestingly, the two views the model found most difficult to correctly differentiate (A4CHLV versus A5CH, and A2CH versus A3CH) were also the two views on which the two experts
disagreed most often. <p>The A4CH view is in an anatomical continuity with the A5CH view. The
difference is whether the scanning plane has been tilted to bring the aortic valve into view, which
would make it A5CH. When the valve is only partially in view, or only in view during part of the
cardiac cycle, the decision becomes a judgement call and there is room for disagreement.
<p>Similarly, the A3CH view differs from the A2CH view only in a rotation of the probe anticlockwise, again to bring the aortic valve into view
It is also interesting to note that the misclassification is not fully asymmetrical. For instance,
while 42 cases of A5CH images are confused with A4CH-LV, there are only 14 occasions of
A4CH-LV images mistaken for A5CH. </h3>
<br> <br>
<img class="img-fluid" src="assets\img\projects\view_classification\fig6.PNG" alt="view classification">
<h3> Fig. 5 t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization of 14 echo views from
the 2-cell-DARTS model </h3>
<br><br> <h3> Figure 6 shows examples of misclassified cases, when the prediction of the 2-cell-DARTS
model disagreed with the expert annotation. The error can be explained by the inherent difficulty
of deciding, even for cardiologist experts, between views that are similar in appearance to human
eyes and are in spatial continuity (case of A4CH/A5CH mix-up), images of poor quality (case of
A4CH/PSAX mix-up), or views in which a same view-defining structure may be present (case of
PSAX-LV/PSAX/AV mix-up).</h3>
<br><br> <h2><b>3. Impact of Image Resolution, Quality, and Dataset Size</b></h2>
<h3> <br><p>The models seem to exhibit a plateau of accuracy between the two larger image resolutions
of 96 × 96 and 128 × 128 pixels (Fig. 7). On the other hand, for the smaller image size of
32 × 32 pixels, the classification performance seems to suffer across all network models, with
a 2.3% to 5.1% reduction in accuracy relative to the resolution of 96 × 96 pixels.<br>
<p>Shown in Fig. 8’s upper panel, is the class-wise view detection accuracy for various input
image resolutions. Notably, not all echo views are affected similarly using lower image resolutions. The drop in overall performance is therefore predominantly caused by a marked decrease
in detection accuracy of only certain views. For instance, A4CH-RV suffers a sharp reduction of
>10% in prediction accuracy when dealing with images of 32 × 32 pixels.<br>
<br><br><br><p>Figure 8’s lower panel shows the relative confusion matrix, illustrating the improvement
associated with using image resolution of 96 × 96 versus 32 × 32 pixels. Being already a difficult view to detect even in higher resolution images, A5CH will have 47 more cases of misclassified images when using images of 32 × 32 pixels. Overall, apical views seem to suffer the
most from lower resolution images, being mainly misclassified as other apical views.<br><p> For instance, the two classes associated with the A4CH will primarily be mistaken for one another.
This is likely because, with a decreased resolution, the details of their distinct features would be
less discernible by the network. Conversely, parasternal views seem to be less affected, and still
detectable in downsampled images. This could be owing to the fact that the relevant features,
on which the model relies for identifying this view, are still present and visible to the model.
Overall, and for almost all echo views, the image size of 96 × 96 pixels appeared to be a good
compromise between classification accuracy and computational costs. <br>
<p>To examine the influence of the size of the training dataset on the model’s performance, we
conducted an additional experiment where we split the training data into sub-datasets with strict
inclusion relationship (i.e., having the current sub-dataset a strict subset of the next sub-dataset),
and ensured all the sub-datasets were consistent (i.e., having the same ratio for each echo view as
in the original training dataset). We then retrained all targeted neural networks on these subdatasets from scratch, and investigated how their accuracy varied with respect to the size of the
dataset used for training the model. The size of the validation and testing datasets, however,
remained unchanged. <br>
<br><br><br> <p> Figure 9 shows a drop in the classification accuracy across all models when smaller sizes of
training data are used for training the networks. However, various models are impacted differently. Suffering from redundancy, deeper neural networks require more training data to achieve
similar performances. DenseNet, with the largest number of trainable parameters, appears to be the one which suffers the most, with a 20% reduction in its classification accuracy, when only 8%
of the training dataset is used. <br>
<p>However, the DARTS-based models appear to be relatively less profoundly affected by the
size of the training dataset, where both models demonstrate no more than 8% drop in their prediction accuracy when deprived of the full training dataset. When using fewer than 12,400
images (i.e., 50% of the training dataset), both DARTS-based models exhibit superior performance over the deeper networks.
Additionally, we hypothesized that the more numerous the echo view classes, the more
difficult the task of distinguishing the views for deep learning models, e.g., because of more
chances of misclassifications among classes. <br> <br><br>This is potentially the underlying reason for the
inconsistent accuracies (84% to 97%) reported in the literature when classifying between
6 to 12 different view classes. To investigate this premise, we considered cases when only
5 or 7 different echo views were present in the dataset. To this end, rather than reducing the
number of classes by merging several views to create new classes which may not be clinically
very helpful, we were selective in choosing some of the existing classes.</h3>
<img class="img-fluid" src="assets\img\projects\view_classification\table2.PNG" alt="table2">
<h3> Table 2 The dependence of overall accuracy on the number of echo views </h3>
<img class="img-fluid" src="assets\img\projects\view_classification\img quality.PNG" alt="img quality">
<h3> Fig 10 Correlation between the classification accuracy and the image quality </h3>
</div>
</div>
</div>
<!-- ======= Project Results ======= -->
<div class="col-12 col-lg-6 col-md-6">
<div class="text-justify">
<div align="justify">
<h2><b>Results and Discussion</h2>
<h2>1. Architecture Search:</b></h2>
<h3>The search took ∼6, 23, 42, and 92 h for image sizes of 32 × 32, 64 × 64, 96 × 96, and
128 × 128 pixels, respectively, on the computing infrastructure described earlier.
Figure 3 shows the best convolutional normal and reduction cells obtained for the input image
size of 128 × 128 pixels. The retained operations were 3 × 3 and 5 × 5 dilated convolutions,
3 × 3 max-pooling, and skip-connection. Each cell is assumed to have two inputs which are
the outputs from the previous and penultimate cells. The output of the cell is defined as the
depth-wise concatenation of all nodes in the cell.
<p>Two network architectures were assembled from the optimal cell; “1-cell-DARTS” composed of one cell only,
and “2-cell-DARTS” formed from a sequential stack of two cells.
Addition of more cells to the network architecture did not significantly improve the prediction
accuracy, as reported in the next section, but increased the number of trainable parameters in the
model and thus the inference time for view classification. Therefore, the models with more than
two cells, i.e., architectures with redundancy, were judged as being comparatively inefficient and
thus discarded. Figure 3 (left side) also shows the full architecture for the 2-cell-DARTS model
for the input image size of 128 × 128 pixels. <br> <br></h3>
<img class="img-fluid" src="assets\img\projects\view_classification\table1.PNG" alt="Arch search">
<h3> Table 1 Experimental results on the test dataset for input sizes of (32 × 32), (64 × 64), (96 × 96),
and (128 × 128) and different network topologies </h3>
<img class="img-fluid" src="assets\img\projects\view_classification\confusion matrix2.PNG" alt="Conf matrix">
<h3> Fig. 4 Confusion matrix for the 2-cell-DARTS model and input image resolution of
128 × 128 pixels. </h3>
<br><h3> On the other hand, echo views with distinct characteristics are easier for the model to
distinguish. For instance, PLAX-full and Suprasternal seem to have higher rates of correct
identification, and the network is confused only on one occasion between these two views.
This is also evident on the t-distributed stochastic neighbor embedding (t-SNE) plot in Fig. 5,
which displays a planar representation of the internal high-dimensional organization of the 14
trained echo view classes within the network’s final hidden layer (i.e., input data of the fully
connected layer). Each point in the t-SNE plot represents an echo image from the test dataset.
<p><br>Noticeably, not only has the network grouped similar images together (a cluster for each
view, displayed with different color), but it has also grouped similar views together (highlighted
with a unique background color). For instance, it has placed A5CH (blue) next to A4CH (dark
brown), and indeed there is some “interdigitation” of such cases, e.g., for those whose classification between A4CH and A5CH might be debatable. Similarly, at the top right, the network
has discovered that the features of the Subcostal-IVC images (green) are similar to the Subcostal
images (red). This shows that the network can point to relationships and organizational patterns
efficiently.</h3>
<div align = "center">
<img class="img-fluid" src="assets\img\projects\view_classification\Misclassify.PNG" alt="misclassified">
<h3> Fig. 6 Three different misclassified examples predicted by the 2-cell-DARTS model for the image
resolution of 128 × 128 pixels</h3>
<img class="img-fluid" src="assets\img\projects\view_classification\fig7.PNG" alt="comparision" align = "center">
<h3> Fig. 7 Comparison of accuracy for different classification models and different image resolutions</h3>
<img class="img-fluid" src="assets\img\projects\view_classification\accuracy.PNG" alt="accuracy">
<h3> Fig. 8 Accuracy of the 2-cell-DARTS model for various input image resolutions</h3>
<img class="img-fluid" src="assets\img\projects\view_classification\comp accuracy.PNG" alt="comp accuracy">
<h3> Fig. 9 Comparison of accuracy of different classification models for image size</h3>
</div>
<br> <h3>For each study,we aimed at including views representing anatomically adjacent or similar imaging planes such
as apical windows (thus challenging for the models to distinguish), as well as other echo
windows. The list of echo views included in each study is provided in Table 2.
The results show an increase in the overall prediction accuracy for the two DARTS-based
models, when given the task of detecting fewer echo view classes and despite having relatively
smaller training datasets to learn from. The 1-cell-DARTS model shows 8% improvement in its
performance when the number of echo views is reduced from 14 to 5. The 2-cell-DARTS model
reaches a maximum accuracy of 99.3%, i.e., higher than any previously reported accuracies for
echo view classification. This highlights the fact that for a direct comparison of the classification
accuracy between the models reported in literature, the number of different echo windows
included in the study must be taken into account. <br>
<p><br> Finally, in order to study the impact of image quality on the classification performance, we
asked a second expert cardiologist to provide an assessment of image quality in the A4CH-LV
views, and assign a quality label to each image where the quality was classified into 5 grades:
very poor, poor, average, good, and excellent. Figure 10 shows the relationship between the
classification accuracy of the 2-cell-DARTS model and the image quality in the test dataset.
The area of the bubbles represents the relative frequency of the images in that quality score
category, with the good category as the dominant grade. This is likely because the image acquisition had been performed mainly by experienced echocardiographers.<br>
<br>The correlation between the classification accuracy and the image quality is evident (p-value
of 0.01). Images labeled as having “excellent” quality, indicated the highest classification accuracy of ∼100%. It is apparent that the discrepancy between the model’s prediction and the expert
annotation is higher in poor quality images. This could potentially be due to the fact that poorly
visible chambers with a low degree of endocardial border delineation could result in some views
being mistaken for other apical windows.
</div>
</div>
</div>
</div>
</div>
</div>
<!-- ======= Project Limitations and conclusion ======= -->
<!-- <div class = "project-Limitations and Conclusion">
<div class="container">
<div class="row justify-content-left">
<div class="col-12">
<div align="justify">
<hr>
<h2><b>4. Study Limitations and Future Work</b></h2>
<h3><p class="text-justify">TThis study sheds light on several possible directions for future work. Herein, we have focused on
the rapid and accurate classification of individual frames from an echo cine loop. Such a task will
be crucial for a real-time view detection system in clinical scenarios where images need to be
processed while they are acquired from the patient and/or where the system is to be used for
operator guidance. However, for offline studies and when the entire cine loop is available, classification of the echo videos could also be of practical use. Some studies have attempted video
classification using the majority vote on some or all frames from a given video.6,34 However, this
approach does not use the temporal information available in the cine loop, such as the movement
of structures during the cardiac cycle. Therefore, a future study could look into using all available
information for view detection. <br>
<br>Our study investigated 2D echocardiography as the clinically relevant modality. Currently,
3D echocardiography suffers from a considerable reduction in frame rate and image quality, and
this has limited its adoption into routine practice over the past decade.46–48 When such issues are
resolved, automatic processing of the 3D modality could also be explored. In the meantime,
2D echocardiography remains unrivalled, particularly when high frame rates are needed.
We investigated the impact of image quality on the classification accuracy for apical fourchamber views only. A more comprehensive examination of the image quality and its influence
on the detection of different echo views would be informative.<br>
<br>The dataset used in this study comprised images acquired using ultrasound equipment from
GE and Philips manufacturers. Although the proposed models do not make any a priori assumptions on data obtained from specific vendors and therefore should be vendor-neutral, echo studies using more diverse ultrasound equipment should still be explored.
Similar to all previous studies, our dataset originated from one medical center, i.e., Imperial
College Healthcare NHS Trust’s echocardiogram database. Representative multi-center patient
data will be essential for ensuring that the developed models will scale up well to other sites and
environments. <br>
Interpreting the results of the proposed models alongside other proposed architectures in the
literature (with a wide range of reported accuracies) was not feasible. This is due to the fact that a
direct comparison of the classification accuracy would require access to the same patient dataset.
At present, no echocardiography dataset and corresponding annotations for view detection are
publicly available.<br>
In order to address such broadly acknowledged shortcomings in the application of deep learning to echocardiography, we are now developing Unity (data.unityimaging.net), a UK collaborative of cardiologists, physiologists and computer scientists, under the aegis of the British
Society of Echocardiography. An image analysis interface has been developed in the form
of a web-based, interactive, real-time platform to capture carefully curated expert annotations
from numerous echo specialists, with patient data provided by over a dozen sites across the UK,
thus ensuring coverage of multiple vendors, systems and environments. All developed models
designed using this annotation biobank (e.g., automated cardiac phase detection,49 left ventricular segmentation,50 and view classification in current study), will be made available under opensource agreements on intsav.github.io.</h3>
<br> <br>
<h2><b>5. Conclusion</b></h2>
<h3> In this study, efficient CNN architectures are proposed for automated identification of the 2D
echocardiographic views. The DARTS method was used in designing optimized architectures for
rapid inference while maintaining high accuracy. A dataset of 14 different echocardiographic
views was used for training and testing the proposed models. Compared with the standard classification CNN architectures, the proposed models are faster and achieve comparable classification performance. Such models can thus be used for real-time detection of the standard
echo views. <br><br>
The impact of image quality and size of the training dataset on the efficacy of the models was
also investigated. Deeper neural network models, with a large number of redundant trainable parameters, require more training data to achieve similar performances. A direct correlation
between the image quality of classification accuracy was observed.
The number of different echo views to be detected has a direct impact on the performance of
the deep learning models, and must be taken into account for a fair comparison of classification
models.<br> <br>
Aggressively downsampled images will result in losing relevant features, thus lowering the
prediction accuracy. On the other hand, while much larger images may be favored for some finegrained applications (e.g., segmentation), their use for echo view classification would offer only
slight improvements in performance (if any) at the expense of more processing and memory
requirements.
</h3>
</div>
</div>
</div>
</div>
</div> -->
<div class="container">
<hr>
<div class="row justify-content-left">
<div class="col-12">
<div class="text-left">
<h1> <b> Project Team </b></h1>
<h6> <a href="https://www.uwl.ac.uk/staff/neda-azarmehr" target="_blank" rel="noopener noreferrer"> Neda Azarmehr </a></h6>
<h6> <a href="https://elisabethlane.github.io/" target="_blank" rel="noopener noreferrer"> Elisabeth S Lane </a></h6>
<h6> <a href="https://twitter.com/intsav_?lang=en-gb" target="_blank" rel="noopener noreferrer"> Robert Labs </a></h6>
<h6> <a href="https://www.uwl.ac.uk/staff/massoud-zolgharni" target="_blank" rel="noopener noreferrer"> Massoud Zolgharni </a></h6>
</div >
</div>
</div>
</div>
<div class="container">
<hr>
<div class="row justify-content-left">
<div class="col-12">
<div class="text-left">
<h1> <b> References</b> </h1>
<br>
<h3> <a href="https://doi.org/10.1117/1.JMI.8.3.034002" >Neural Architecture Search of Echocardiography View Classifiers</a></h3>
</div>
</div>
</div>
</div>
</section><!-- End project-body -->
</main><!-- End #main -->
<!-- ======= Footer - DONT CHANGE ======= -->
<footer id="footer">
<div class="container d-md-flex py-4">
<div class="me-md-auto text-center text-md-start">
<div class="copyright">
© Copyright <strong><span>IntSaV</span></strong>. All Rights Reserved
</div>
</div>
<div class="social-links text-center text-md-right pt-3 pt-md-0">
<a href="#" class="twitter"><i class="bx bxl-twitter"></i></a>
<a href="#" class="facebook"><i class="bx bxl-facebook"></i></a>
<a href="#" class="instagram"><i class="bx bxl-instagram"></i></a>
<a href="#" class="google-plus"><i class="bx bxl-skype"></i></a>
<a href="#" class="linkedin"><i class="bx bxl-linkedin"></i></a>
</div>
</div>
</footer><!-- End Footer -->
<div id="preloader"></div>
<a href="#" class="back-to-top d-flex align-items-center justify-content-center"><i class="bi bi-arrow-up-short"></i></a>
<!-- Vendor JS Files -->
<script src="assets/vendor/aos/aos.js"></script>
<script src="assets/vendor/bootstrap/js/bootstrap.bundle.min.js"></script>
<script src="assets/vendor/glightbox/js/glightbox.min.js"></script>
<script src="assets/vendor/isotope-layout/isotope.pkgd.min.js"></script>
<script src="assets/vendor/php-email-form/validate.js"></script>
<script src="assets/vendor/purecounter/purecounter.js"></script>
<script src="assets/vendor/swiper/swiper-bundle.min.js"></script>
<!-- Template Main JS File -->
<script src="assets/js/main.js"></script>
</body>
</html>