Week #17

18 Feb 2019

Divide the axial planes slices into 18 groups

Fig. 1 Directory illustration

Data in slices

Table. 1 Slices info for axial plane

Slice num	Class Num	Class labels
0	167	0-165,167,168,169 except 86,91
1	172	0-171,177 no 91
2	174	0-175,176 except 91
3	174	0-173,176 except 91
4	175	0-171,174,176,177,180 except 91
5	171	0-169,174,177 except 91
6	170	0-170 except 91
7	168	0-169 except 91
8	153	0-151,155,156 except 91
9	145	0-144,146 except 91
10	135	0-135,137 except 91,131
11	129	0-127,129,130 except 91
12	125	0-124,126 no 91
13	123	0-120, 122, 126, 128 no 91
14	117	0-116,118, no 91
15	97	0-102,104,109 no 32,34,36,37,38,40,41,91
16	1	0
17	1	0
18	1	0
total	178	0-177,180, no 91

From the table above, we observe that there are in total 178 different class labels for axial plane, so the network is defined to have 178 classes as output.
Training on slice0, where there are 3330 images in total, with 3,303,618 the loss could reach 0.007 within 35 epochs. The loss is represented in the figure below.

Fig. 2 Loss for slice0 Fig. 2 Loss for slice1 Fig. 3 Loss for slice2 after training with slice1

Parc_1 experiments

Slice Idx	training subjects	slice per subjects	total images	image size	learing rate	optimization	Paremeters	Batch size	epoch	minimum loss reached
0	330	10	3300	128x128	0.001	Adam	3,303,618	5	35	0.007
1	330	10	3300	128x128	0.001	Adam	3,303,618	5	50	0.066
1,2	330	10	6600	128x128	0.001	Adam	3,303,618	5	50	0.231
2	330	10	3300	128x128	0.001	Adam	3,303,618	50	50	0.242
3	330	10	3300	128x128	0.001	Adam	3,303,618	10	50	0.533

I have done an experiments using these slide group by group. Batch size=50. parameter size 3,303,618 Here are the results:

slicegroup	epoch0 loss	epoch50 loss
0	0.802	0.008
1	0.121	0.069
2	0.361	0.235
3	0.729	0.540
4	1.001	0.771
5	1.363	1.088
6	1.516	1.287
7	1.415	1.234
8	1.274	1.099
9	1.107	0.939
10	1.094	0.936
11	0.956	0.832
12	0.684	0.584
13	0.339	0.256
14	0.063	0.032

Decrease learning rate every 30 epochs: