-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathOHBM2024.html
941 lines (838 loc) · 44.7 KB
/
OHBM2024.html
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
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
<link href="css/fontawesome-free-6.2.1-web//css/all.css" rel="stylesheet">
<script src="lib/colorbrewer.v1.min.js" charset="utf-8"></script>
<script src="lib/colorStringStandalone.js" charset="utf-8"></script>
<script type="text/javascript" src="lib/jquery-2.2.4.min.js"></script>
<script src="https://cdn.jsdelivr.net/npm/niivue@latest/dist/niivue.js"></script>
<!-- Load d3.js -->
<title>OHBM 2024 lecture</title>
<meta name="description" content="OHBM Educational course 'Deep Learning in Neuroimaging'">
<meta name="author" content="Sergey M Plis">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="black-translucent">
<link rel="stylesheet" href="dist/reset.css">
<link rel="stylesheet" href="dist/reveal.css">
<!-- Code syntax highlighting -->
<link rel="stylesheet" href="plugin/highlight/monokai.css" id="highlight-theme">
<!-- <link rel="stylesheet" href="lib/css/zenburn.css"> -->
<link rel="stylesheet" href="css/custom.css">
<link rel="stylesheet" href="dist/theme/aml.css" id="theme">
<!-- Printing and PDF exports -->
<script>
var link = document.createElement( 'link' );
link.rel = 'stylesheet';
link.type = 'text/css';
link.href = window.location.search.match( /print-pdf/gi ) ? 'css/print/pdf.css' : 'css/print/paper.scss';
document.getElementsByTagName( 'head' )[0].appendChild( link );
</script>
</head>
<body>
<div class="reveal">
<div class="slides">
<section>
<section data-state="brightside">
<p>
<h2>Supervised and Self-supervised Deep Learning for Structural <span class="fragment fade-out">and Functional </span>Neuroimaging</h2>
<h3>Sergey Plis</h3>
<h3 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1">OHBM 2024: Deep learning in Neuroimaging</h3>
<img style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255,
0.8);" width="300"
src="figures/OHBM2024banner.png"
alt="IJCNN">
<img style="border:0;
box-shadow: 0px
0px 0px;"
width="300"
src="figures/TReNDS_logo_light.png"
alt="logo">
<img style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255,
0.8);" height="100"
src="figures/neuroneural_logo.png"
alt="Neuroneural">
</section>
<section>
<h3>Outline for the lecture</h3>
<ul>
<li class="fragment roll-in"> A Short and Based History of CNNs
<li class="fragment roll-in"> Building Blocks of CNNs
<li class="fragment roll-in"> Predictive Model
<li class="fragment roll-in"> Semantic Segmentation and More Blocks
<li class="fragment roll-in"> U-Net - a Powerful Segmentation Model
<li class="fragment roll-in"> MeshNet - Still Powerful but tiny
<li class="fragment roll-in"> Self-Supervised, aka Synthetic Data, Training
</ul>
</section>
</section>
<section>
<section data-vertical-align-top>
<h2 style="margin-top: -80px;">STRUCTURAL DATA</h2>
<iframe id="gl" src="niivue.html" style="width:100%; height:700px;" frameborder="0" allowfullscreen></iframe>
</section>
<section>
<h2>Convolutional Neural Networks</h2>
<h2>work great for structure</h2>
</section>
</section>
<section>
<section>
<h2> a very brief and biased</h2>
<h1>history</h1>
</section>
<section data-vertical-align-top>
<div id="header-right" style="margin-right: -100px;">
<img width="230"
src="figures/HubelWiesel.jpg" alt="hubelWiesel">
<div style="font-size:14pt;">David Hubel</div>
<div style="font-size:14pt;">Torsten Wiesel</div>
</div>
<h2>Cat's brain <alert>1962</alert></h2>
<img width="70%"
src="figures/cats_brain.png" alt="cat's brain">
<div class="slide-footer">
<a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC1359523/pdf/jphysiol01247-0121.pdf">Receptive fields, binocular interaction and functional architecture in the cat's visual cortex</a>
</div>
</section>
<section>
<div id="header-right" style="margin-right: -100px;">
<img width="130"
src="figures/fukushima.jpg" alt="fukushima">
<div style="font-size:14pt;">Kunihiko Fukushima</div>
</div>
<h2>Neurocognitron <alert>1979</alert></h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); margin-top: -20px;" width="900"
src="figures/neurocognitron_.png" alt="neurocognitron">
<div class="slide-footer">
<a href="https://www.cs.princeton.edu/courses/archive/spr08/cos598B/Readings/Fukushima1980.pdf">Neocognitron - A Self-organizing Neural Network Model for a Mechanism of Pattern Recognition Unaffected by Shift in Position</a>
</div>
</section>
<section>
<div id="header-right" style="margin-right: -100px;">
<img width="130"
src="figures/waibel.jpg" alt="waibel">
<div style="font-size:14pt;">Alexander Weibel</div>
</div>
<h2>Time Delay Neural Network <alert>1989</alert></h2>
<img style="margin-top: -20px;" width="40%"
src="figures/tdnn_.png" alt="tdnn">
<div class="slide-footer">
<a href="https://www.cs.toronto.edu/~fritz/absps/waibelTDNN.pdf">Phoneme Recognition Using Time Delay Neural Networks</a>
</div>
<aside class="notes">
Note 1D channels!
</aside>
</section>
<section data-vertical-align-top>
<div id="header-right">
<img width="130"
src="figures/lecun.png" alt="lecun">
<div style="font-size:14pt;">Yann LeCun</div>
</div>
<h2 style="margin-top: -50px;">CNN <alert>1989</alert></h2>
<img style="margin-top: -50px;" width="73%"
src="figures/cnn.png" alt="cnn">
<div class="slide-footer">
<a href="http://yann.lecun.org/exdb/publis/pdf/lecun-89e.pdf">Backpropagation Applied to Handwritten Zip Code Recognition</a>
</div>
<aside class="notes">
Note undersampling to reduce layers!
</aside>
</section>
<section>
<div id="header-right">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); " width="130"
src="figures/lecun.png" alt="lecun">
<div style="font-size:14pt;">Yann LeCun</div>
</div>
<h2>CNN <alert>1998</alert></h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/lenet.png" alt="lenet">
<div class="slide-footer">
<a href="http://yann.lecun.com/exdb/publis/pdf/lecun-01a.pdf">Gradient-Based Learning applied to Document Recognition</a>
</div>
<aside class="notes">
Note subsampling! No pooling yet!
</aside>
</section>
</section>
<section>
<section>
<h1>Building blocks I</h1>
</section>
<section>
<h2>to build CNNs we need:</h2>
<h1 class="fragment roll-in" >Convolutions</h1>
<h1 class="fragment roll-in" >Max Pooling</h1>
<h1 class="fragment roll-in" >Skip Connections</h1>
</section>
<section>
<h2>Convolving a kernel with an image</h2>
<row>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_convo1.gif" alt="convolution">
\[
\left(
\begin{array}{ccc}
0 & 1 & 2 \\
2 & 2 & 0 \\
0 & 1 & 2 \\
\end{array}
\right)
\]
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_convo2.gif" alt="convolution 2">
</col50>
</row>
<div class="slide-footer" style="margin-bottom: -50px;">
<a href="https://arxiv.org/pdf/1603.07285.pdf">A guide to convolution arithmetic for deep learning</a>
</div>
</section>
<section>
<h2>Convolution: frame by frame</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_flat_conv.svg" alt="convolution">
</section>
<section>
<h2>Pooling: maxpooling</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_maxpooling.gif" alt="max pooling">
</section>
<section>
<h2>Pooling: maxpooling</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_flat_maxpooling.svg" alt="max pooling">
</section>
<section data-vertical-align-top>
<h2 style="margin-top: -60px;">What happens in 3D?</h2>
<img class="fragment roll-in" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8); justify-content: center;"
width="80%"
src="figures/volumetric_convolution.svg"
alt="Volumetric Convolution">
<img class="fragment roll-in" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8); justify-content: center;"
width="100%"
src="figures/volumetric_convolution_layer.svg"
alt="Volumetric Convolution Layer">
</section>
<section>
<h1>skip connections</h1>
</section>
<section>
<h3>Highway networks (May 2015 on arxiv)</h2>
<ul style="list-style-type: none; font-size: 32px;">
<li class="fragment roll-in">
$$
\vec{y} = H(\vec{x}, \bm{W}_H)
$$
<li class="fragment roll-in">
$$
\vec{y} = H(\vec{x}, \bm{W}_H) \odot T(\vec{x}, \bm{W}_T) + \vec{x} \odot C(\vec{x}, \bm{W}_C)
$$
<li class="fragment roll-in">
$$
\vec{y} = H(\vec{x}, \bm{W}_H) \odot T(\vec{x}, \bm{W}_T) + \vec{x} \odot (1 - T(\vec{x}, \bm{W}_T))
$$
<li class="fragment roll-in">
$$
\vec{y} =
\left\{
\begin{array}{ll}
\vec{x} & \mbox{if }\;\;T(\vec{x}, \bm{W}_T)=0,\\
H(\vec{x}, \bm{W}_H) & \mbox{if }\;\;T(\vec{x}, \bm{W}_T)=1
\end{array}
\right.
$$
</ul>
<div class="slide-footer" style="text-align: left;">
<a href="https://arxiv.org/abs/1505.00387">Highway Networks</a><br>
<a href="https://arxiv.org/abs/1507.06228">Training Very Deep Networks</a>
</div>
<aside class="notes">
In 2015 Rupesh Srivastawa, Klaus Greff and Yurgen Schmidhuber made the following, a bit elaborate, observation. If we embellish the regular layer in a feed forward neural network with a transfer and carry gates, we will enable gradients flow along the entire depth of a deep network.<br>
Let's see how they have done it.
</aside>
</section>
<section data-fullscreen>
<alert>Train models with 100 of layers instead of just 10 before</alert>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_HW_performance.png" alt="HWN">
<aside class="notes">
The networks that they have trained using traditional layers and highway layers were of depth 10, 20, 50, and a 100. Plain networks would train increasingly poorly with depth, while highway networks retained ability to train. Importantly the results were comparable with the ones of fitnets (trained by mimicing a teacher).
</aside>
</section>
<section>
<h2>Residual Networks (block)</h2>
<row>
<col50>
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_resnets_block.svg" alt="resnet block">
</col50>
</row>
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a> December 2015 on arxiv
</div>
</section>
<section>
<h2>Residual Networks (full)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_resnets_whole.svg" alt="resnet whole">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a>
</div>
</section>
<section>
<h2>Residual Networks (performance)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1500"
src="figures/CNN_resnets_perf.svg" alt="resnet perf">
<div class="slide-footer">
<a href="https://arxiv.org/pdf/1512.03385.pdf">Deep Residual Learning for Image Recognition</a>
</div>
</section>
<section>
<h2>Error surface effect of skip connection</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/skip_smooth.png" alt="landscape">
<div class="slide-footer">
<a href="https://papers.nips.cc/paper/7875-visualizing-the-loss-landscape-of-neural-nets.pdf">Visualizing the Loss Landscape of Neural Nets</a>
</div>
<aside class="notes">
</aside>
</section>
<section>
<h2>Dense Networks (architecture)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="80%"
src="figures/CNN_densenet.svg" alt="dense diagram">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1608.06993">Densely Connected Convolutional Networks</a>
</div>
<aside class="notes">
</aside>
</section>
<section>
<h2>Dense Networks (effect)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/densenet.png" alt="dense">
<div class="slide-footer">
<a href="https://arxiv.org/abs/1608.06993">Densely Connected Convolutional Networks</a>
</div>
<aside class="notes">
</aside>
</section>
</section>
<section>
<section>
<h2>Volumetric Predictive Model</h2>
<img style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8); justify-content: center;"
width="120%"
src="figures/Anees_ResNetModel.png"
alt="Anees' ResNet">
<div class="slide-footer">
<a href="https://www.sciencedirect.com/science/article/abs/pii/S0165027020301242">Deep Residual Learning for Neuroimaging: An application to
Predict Progression to Alzheimer’s Disease </a> 2020
</div>
</section>
<section>
<h2 style="margin-top: -50px;">Alzheimer's Disease prediction</h2>
<img style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8); justify-content: center;"
width="100%"
src="figures/Anees_ADresults.png"
alt="AD vs Cognitively Normal">
<div class="slide-footer">
<a href="https://www.sciencedirect.com/science/article/abs/pii/S0165027020301242">Deep Residual Learning for Neuroimaging: An application to
Predict Progression to Alzheimer’s Disease </a> 2020
</div>
</section>
<section data-background="figures/AneesCommunicationResults_margin.svg" data-background-size="contain" data-vertical-align-top>
<h2 style="margin-top: -80px; margin-right: -400px;">Improves on Big Data</h2>
<div class="slide-footer" style="margin-bottom: -70px;">
<a href="https://www.nature.com/articles/s41467-020-20655-6">Deep learning encodes robust discriminative neuroimaging representations to outperform standard machine learning</a> 2021
</div>
</section>
</section>
<section>
<section data-background="figures/background_egmentation.png" data-background-size="cover" data-vertical-align-top>
<h2 style="margin-top: -80px;">Brain Segmentation</h2>
<img class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="250"
src="figures/CNN_freesurfer.png" alt="david">
<!-- <iframe id="gl" src="niivue.html" style="width:100%; height:700px;" frameborder="0" allowfullscreen></iframe> --> </section>
<section>
<h2>Missing pieces:</h2>
<h1 class="fragment roll-in" >One-Convolution</h1>
<h2 class="fragment roll-in" >Fully convolutional networks</h2>
<h1 class="fragment roll-in" >Up-convolution</h1>
</section>
<section>
<h2>One-convolution</h2>
<row>
<col50>
<img class="fragment roll-in" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8); justify-content: center;"
width="40%"
src="figures/volumetric_1convolution.svg"
alt="Volumetric 1 conv">
</col50>
<col50>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_one_conv.gif" alt="one_conv">
</col50>
</row>
</section>
<section data-fullscreen>
<h3>Replacing feed forward with convolutional</h3>
<img class="fragment roll-in" style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="80%"
src="figures/CNN_semantic_cc_1.png" alt="cc 1">
<img class="fragment roll-in" style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); margin-top: -30px;" width="80%"
src="figures/CNN_semantic_cc_3.png" alt="cc 1">
</section>
<section>
<div id="header-right">
<img width="300"
src="figures/CNN_upconvolution.gif" alt="gif upconv">
</div>
<h2>Upconvolution</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_upconvolution.png" alt="upconv">
<div class="slide-footer">
<a href="https://datascience.stackexchange.com/a/20176">Image from a comment on stackexchange</a>
</div>
</section>
<section data-fullscreen>
<h2>Fully Convolutional Model (2014)</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="1000"
src="figures/CNN_semantic_ff.png" alt="final model">
<div class="slide-footer">
<a style="margin-left: 500px;" href="https://arxiv.org/abs/1411.4038" target="_blank">Fully Convolutional Networks for Semantic Segmentation</a><br>
</div>
</section>
<section>
<h2 style="margin-left: -80px;">Autoencoder: an encoder/decoder model</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); " width="850"
src="figures/AE.svg" alt="Autoencoder">
<div class="slide-footer">
<a href="https://www.cs.toronto.edu/~hinton/science.pdf">Hinton, G.E. and Salakhutdinov, R.R., 2006. Reducing the dimensionality of data with neural networks. science, 313(5786), pp.504-507.</a>
</div>
</section>
</section>
<section>
<section>
<h1>Segmentation Model</h1>
</section>
<section>
<h2>Let's put together</h2>
<h1 class="fragment roll-in" >Upconvolution</h1>
<h2 class="fragment roll-in" >Fully convolutional networks</h2>
<h2 class="fragment roll-in" >Encoder-decoder architecture</h2>
<h1 class="fragment roll-in" >Skip Connections</h1>
</section>
<section data-fullscreen>
<div id="header-left" style="margin-top: 100px; margin-left: 50px;">
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); " width="130"
src="figures/Olaf_Ronneberger_Portrait.jpg" alt="Ronneberger">
<div style="font-size:14pt;">Olaf Ronneberger</div>
</div>
<h2>One U-shaped Network</h2>
<h1 class="fragment roll-in">U-Net</h1>
<img width="70%" style="margin-top: -260px;"
src="figures/V-Net.png" alt="final model">
<div class="slide-footer" style="margin-left: 200px;">
<a href="https://arxiv.org/abs/1505.04597" target="_blank">
Ronneberger et al. U-net: Convolutional networks for biomedical image segmentation. MICCAI 2015
</a><br>
<a href="https://arxiv.org/abs/1606.06650" target="_blank">
Çiçek et al. 3D U-Net: learning dense volumetric segmentation from sparse annotation. MICCAI 2016
</a>
</div>
</section>
<section data-background="figures/unet_takeover.png" data-background-size="contain" data-vertical-align-top>
<h2 style="margin-top: -70px; margin-left: -350px;">U-Net took over the world</h2>
</section>
<section>
<h1>But there is a problem</h1>
parameters and memory usage
</section>
<section data-fullscreen>
<h2>Consider a small U-net</h2>
<row>
<col50>
<img class="fragment roll-in" class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="100%"
src="figures/unet_arc.png" alt="small U-Net">
</col50>
<col50>
<img class="fragment roll-in" class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="100%"
src="figures/unet_table.png" alt="U-Net parameter table">
</col50>
</row>
</section>
<section data-fullscreen>
<h2>Skip-connections need memory</h2>
<img width="70%" style="margin-top: -70px;"
src="figures/V-Net.png" alt="final model">
</section>
<section data-fullscreen>
<h2>We could train on subcubes</h2>
<img width="70%" style="margin-top: -70px;"
src="figures/subvolume_prediction_pipeline.png" alt="final model">
</section>
<section>
<h3>Full volume training and inference gives better accuracy</h3>
<h3>but requires 80Gb GPU for training and more than 10Gb for inference</h3>
</section>
<section>
<h3>Mechanisms for Sharing Neuroimaging AI Models</h3>
<ul>
<li class="fragment toll-in"> Native Installation (conda, CUDA, and all that Jazz)
<li class="fragment toll-in"> Containers <i class="fa-solid fa-box-open"></i> (see <a href="https://github.com/neuronets/nobrainer">nobrainer</a>)
<li class="fragment toll-in"> Running Models in the Cloud
<li class="fragment toll-in"> Edge-Based Deployment
</ul>
</section>
</section>
<section>
<section>
<h1>A sparse solution</h1>
</section>
<section>
<h2>enter Dilated convolution</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="60%"
src="figures/CNN_dilation.gif" alt="dilated">
</section>
<section>
<h2><span class="fragment fade-out" data-fragment-index="4">U-Net</span> building blocks</h2>
<ul>
<li class="fragment fade-out" data-fragment-index="0">Upconvolution
<li>Fully convolutional networks
<li class="fragment fade-out" data-fragment-index="1">Encoder-decoder architecture
<li class="fragment fade-out" data-fragment-index="2">Skip Connections
<li class="fragment roll-in" data-fragment-index="3">Dilated Convolutions
</ul>
</section>
<section>
<div id="header-left" style="margin-left: -200px; z-index: 1500;">
<img width="200px" style="margin-bottom: -5%;"
src="figures/AlexFedorov.jpg" alt="Alex"><br>
<small>Alex Fedorov</small>
</div>
<h2>MeshNet</h2>
<row>
<col40>
<img class="fragment roll-in" class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="100%"
src="figures/meshnet_arch.png" alt="MeshNet">
</col40>
<col60>
<img class="fragment roll-in" class="stretch" style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255, 0.8);" width="100%"
src="figures/mntable.png" alt="MeshNet parameter table">
<ul style="font-size:30px;">
<li class="fragment roll-in">72516 vs 23523355 U-Net parameters
<li class="fragment roll-in">but that's for 21 channel MeshNet
<li class="fragment roll-in">5 channel MeshNet is about 4x smaller
</ul>
</col60>
</row>
<div class='slide-footer' style="text-align: left;">
Fedorov et al. End-to-end learning of brain tissue segmentation
from imperfect labeling. IJCNN 2017
</div>
</section>
<section data-vertical-align-top data-background-iframe="https://brainchop.org"
data-background-interactive>
<h3 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1">Now we can deliver models to users of any skill</h3>
<div class='slide-footer'>
<a href="https://brainchop.org">brainchop.org</a><br>
<a href="https://github.com/neuroneural/brainchop">https://github.com/neuroneural/brainchop <i class="fab fa-github"></i></a>
</div>
</section>
</section>
<section>
<section>
<h1>Self-Supervised</h1>
<h2>aka Synthetic Data training</h2>
</section>
<section data-background="figures/forging_plows.jpg" data-background-size="cover">
<h3 style="margin-left: -200px;">Do we need all data in the world to train robust models?</h3>
</section>
<section>
<h3>data generation</h3>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="100%"
src="figures/Synth_generation_overview.png" alt="Synth Generation">
</section>
<section>
<h2>SynthSeg</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="100%"
src="figures/SynthSeg_segmentations.png" alt="Synth Example">
<div class='slide-footer'>
<a href="https://github.com/BBillot/SynthSeg">SynthSeg github</a>
</div>
</section>
<section>
<h2>SynthStrip</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="120%"
src="figures/SynthStrip_example.png" alt="Synth Strip Example">
<div class='slide-footer'>
<a href="https://surfer.nmr.mgh.harvard.edu/docs/synthstrip/">SynthStrip page</a>
</div>
</section>
<section>
<h2>SynthSR</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="80%"
src="figures/SynthSR_examples.png" alt="Synth SR Example">
<div class='slide-footer'>
<a href="https://github.com/BBillot/SynthSR">SynthSR github</a>
</div>
</section>
<section>
<h2>The Catch: Training takes a week</h2>
GPU utilization
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="100%"
src="figures/SynthSeg_utilization.png" alt="synthseg training">
</section>
<section>
<h2>Decouple training and generation</h2>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1);" width="60%"
src="figures/differential_wirehead.gif" alt="differential">
</section>
<section>
<h3>really Decouple training and generation</h3>
<img style="border:0; box-shadow: 0px 0px 0px rgba(150, 150, 255, 1); margin-top: -50px;" width="100%"
src="figures/wirehead_poster.png" alt="wirehead">
</section>
<section data-vertical-align-top data-background-iframe="https://brainchop.org"
data-background-interactive>
<h3 style="text-shadow: 4px 4px 4px #002b36; color: #93a1a1">Demo the robust model</h3>
<div class='slide-footer'>
<a href="https://brainchop.org">brainchop.org</a><br>
<a href="https://github.com/neuroneural/brainchop">https://github.com/neuroneural/brainchop <i class="fab fa-github"></i></a>
</div>
</section>
</section>
<section>
<section>
<h3>today We have learned</h3>
<ol>
<li class="fragment roll-in"> How volumetric CNNs are built for working with structural MRI
<li class="fragment roll-in"> How U-Net, the most popular semantic segmentation model, works
<li class="fragment roll-in"> How to simplify end-users lives with MeshNet via zero-footprint deep learning
<li class="fragment roll-in"> How to efficiently utilize self-supervised learning to train robust segmentation models
<li class="fragment roll-in"> How to use segmentation models anywhere there is a browser (shameless plug) <a href="http://brainchop.org" target="_blank">brainchop.org</a>
</ol>
</section>
<section data-vertical-align-top data-background-video="figures/brainchop_final_twitter.mp4" data-background-size="cover" data-background-video-loop=true>
<h1>Questions?</h1>
</section>
</section>
</section>
</div>
</div>
<script src="dist/reveal.js"></script>
<link rel="stylesheet" href="plugin/highlight/monokai.css">
<script src="plugin/highlight/highlight.js"></script>
<script src="plugin/math/math.js"></script>
<script src="plugin/chalkboard/plugin.js"></script>
<script src="plugin/notes/notes.js"></script>
<script src="plugin/zoom/zoom.js"></script>
<script src="plugin/fullscreen/fullscreen.js"></script>
<script src="plugin/menu/menu.js"></script>
<script src="plugin/verticator/verticator.js"></script>
<link rel="stylesheet" href="plugin/verticator/verticator.css">
<script>
// Full list of configuration options available at:
// https://github.com/hakimel/reveal.js#configuration
Reveal.initialize({
// history: true,
hash: true,
margin: 0.01,
minScale: 0.01,
maxScale: 1.23,
center: true,
controls: false,
keyboard: true,
overview: true,
transition: 'slide', // Transition style: none/fade/slide/convex/concave/zoom
transitionSpeed: 'slow', // Transition speed: default/fast/slow
menu: {
themes: false,
openSlideNumber: true,
openButton: false,
},
keyboard: {
32: 'next', // Space - next fragment
33: 'prev', // Page Up - previous fragment
34: 'next', // Page Down - next fragment
37: 'left', // Left arrow - previous slide
39: 'right', // Right arrow - next slide
70: 'fullscreen', // 'f' key - fullscreen
},
// Add mouseWheel option to handle clicker events
mouseWheel: true,
// Customize mouse button behavior
navigationMode: 'default',
// Override click behavior
previewLinks: false,
// Enable the slide number display
slideNumber: true,
// Enable the progress bar
progress: true,
chalkboard: {
boardmarkerWidth: 1,
chalkWidth: 2,
chalkEffect: 1,
toggleNotesButton: false,
toggleChalkboardButton: false,
slideWidth: Reveal.width,
slideHeight: Reveal.height,
// src: "chalkboards/chalkboard_em2.json",
readOnly: false,
theme: "blackboard",
eraser: { src: "plugin/chalkboard/img/sponge.png", radius: 30},
},
math: {
mathjax: 'https://cdn.jsdelivr.net/gh/mathjax/[email protected]/MathJax.js',
config: 'TeX-AMS_SVG-full',
// pass other options into `MathJax.Hub.Config()`
TeX: {
Macros: {
RR: '\\mathbb{R}',
PP: '\\mathbb{P}',
EE: '\\mathbb{E}',
NN: '\\mathbb{N}',
vth: '\\vec{\\theta}',
loss: '{\\cal l}',
hclass: '{\\cal H}',
CD: '{\\cal D}',
def: '\\stackrel{\\text{def}}{=}',
pag: ['\\text{pa}_{{\cal G}^{#1}}(#2)}', 2],
vec: ['\\boldsymbol{\\mathbf #1}', 1],
set: [ '\\left\\{#1 \\; : \\; #2\\right\\}', 2 ],
bm: ['\\boldsymbol{\\mathbf #1}', 1],
argmin: ['\\operatorname\{arg\\,min\\,\}'],
argmax: ['\\operatorname\{arg\\,max\\,\}'],
softmax: ['\\operatorname\{soft\\,max\\,\}'],
prob: ["\\mbox{#1$\\left(#2\\right)$}", 2],
},
loader: {load: ['[tex]/color']},
extensions: ["color.js"],
tex: {packages: {'[+]': ['color']}},
svg: {
fontCache: 'global'
}
}
},
plugins: [ Verticator, RevealMath, RevealChalkboard, RevealHighlight, RevealNotes, RevealZoom, RevealMenu ],
});
// Custom key handling for Shift+Space
document.addEventListener('keydown', function(event) {
if (event.shiftKey && event.keyCode === 32) {
event.preventDefault();
Reveal.prev();
}
});
// Custom configuration for mouse buttons
Reveal.configure({
keyboard: {
// Left mouse click (keyCode 1) - next
1: 'next',
// Right mouse click (keyCode 3) - previous
3: function() {
Reveal.prev();
return false; // Prevent context menu
}
}
});
// Prevent context menu on right-click
document.addEventListener('contextmenu', function(e) {
e.preventDefault();
});
Reveal.configure({ fragments: true }); // set false when developing to see everything at once
Reveal.configure({ slideNumber: true });
//Reveal.configure({ history: true });
Reveal.configure({ slideNumber: 'c / t' });
Reveal.addEventListener( 'darkside', function() {
document.getElementById('theme').setAttribute('href','dist/theme/aml_dark.css');
}, false );
Reveal.addEventListener( 'brightside', function() {
document.getElementById('theme').setAttribute('href','dist/theme/aml.css');
}, false );
</script>
<!-- <script type="module" async> -->
<!-- // uses the version of niivue from the main branch of the niivue repository. -->
<!-- // This is the latest development version and may not be stable, and may not reflect -->
<!-- // the functionality of the latest release on NPM. -->
<!-- import * as niivue from "https://niivue.github.io/niivue/dist/index.js" -->
<!-- var volumeList = [ -->
<!-- { url: "figures/t1_crop.nii.gz" }, -->
<!-- ]; -->
<!-- var nv = new niivue.Niivue({ -->
<!-- isResizeCanvas: true, -->
<!-- isOrientationCube: true, -->
<!-- crosshairColor: [0, 1, 0, 0.5], -->
<!-- }); -->
<!-- nv.attachTo("gl"); -->
<!-- nv.loadVolumes(volumeList); -->
<!-- </script> -->
<style type="text/css">
/* 1. Style header/footer <div> so they are positioned as desired. */
#header-left {
position: absolute;
top: 0%;
left: 0%;
}
#header-right {
position: absolute;
top: 0%;
right: 0%;
}
#footer-left {
position: absolute;
bottom: 0%;
left: 0%;
}
</style>
<!-- // 2. Create hidden header/footer -->
<div id="hidden" style="display:none;">
<div id="header">
<div id="header-left"><h4>OHBM2024</h4></div>
<div id="header-right"> <img style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255,
0.8);" width="100"
src="figures/neuroneural_logo.png"
alt="Neuroneural"></div>
<div id="footer-left">
<img style="border:0; box-shadow: 0px 0px 0px
rgba(150, 150, 255,0.8);"
width="200"
src="figures/OHBM2024.png"
alt="OHBM">
</div>
</div>
</div>
<script type="text/javascript">
// 3. On Reveal.js ready event, copy header/footer <div> into each `.slide-background` <div>
var header = $('#header').html();
if ( window.location.search.match( /print-pdf/gi ) ) {
Reveal.addEventListener( 'ready', function( event ) {
$('.slide-background').append(header);
});
}
else {
$('div.reveal').append(header);
}
</script>
</body>
</html>