Useful statistics for training models #20
Description
I computed some stats about the dataset which could be considered for training the segmentation models
Subjects with and without lesions
It is useful to know the exact number of either of them so as to think of a curriculum learning strategy involving training the model for a few "warm-up" epochs only on the subjects with lesions and gradually introducing subjects who do not.
Number of subjects without lesion: 56
['sub-m969884', 'sub-m139339', 'sub-m456943', 'sub-m831195', 'sub-m376420', 'sub-m824387', 'sub-m116619', 'sub-m598270', 'sub-m769014', 'sub-m804794', 'sub-m362157', 'sub-m927055',
'sub-m665521', 'sub-m906416', 'sub-m162129', 'sub-m709160', 'sub-m852614', 'sub-m902962', 'sub-m659205', 'sub-m843987', 'sub-m128628', 'sub-m884947', 'sub-m012474', 'sub-m053662',
'sub-m998939', 'sub-m373162', 'sub-m711452', 'sub-m073580', 'sub-m380212', 'sub-m597981', 'sub-m116500', 'sub-m790028', 'sub-m300747', 'sub-m991840', 'sub-m987382', 'sub-m936588',
'sub-m747612', 'sub-m854598', 'sub-m838132', 'sub-m431499', 'sub-m387058', 'sub-m737478', 'sub-m090343', 'sub-m627960', 'sub-m441629', 'sub-m339071', 'sub-m206271', 'sub-m550628',
'sub-m472036', 'sub-m553941', 'sub-m358902', 'sub-m826180', 'sub-m491476', 'sub-m554105', 'sub-m919335', 'sub-m299563']
Number of subjects with lesion: 163
['sub-m703984', 'sub-m545591', 'sub-m052556', 'sub-m552033', 'sub-m425924', 'sub-m531168', 'sub-m508874', 'sub-m205815', 'sub-m990877', 'sub-m484245', 'sub-m757346', 'sub-m868134',
'sub-m723132', 'sub-m738530', 'sub-m322775', 'sub-m362600', 'sub-m707812', 'sub-m463857', 'sub-m597865', 'sub-m378204', 'sub-m026506', 'sub-m818513', 'sub-m718495', 'sub-m572861',
'sub-m563712', 'sub-m977362', 'sub-m978163', 'sub-m829931', 'sub-m991145', 'sub-m295736', 'sub-m159764', 'sub-m531317', 'sub-m158425', 'sub-m360832', 'sub-m243433', 'sub-m142435',
'sub-m221398', 'sub-m762797', 'sub-m724575', 'sub-m786260', 'sub-m560928', 'sub-m275415', 'sub-m818091', 'sub-m808926', 'sub-m522051', 'sub-m117189', 'sub-m556439', 'sub-m774069',
'sub-m220491', 'sub-m434248', 'sub-m916671', 'sub-m694074', 'sub-m222399', 'sub-m839135', 'sub-m350871', 'sub-m763939', 'sub-m739531', 'sub-m793289', 'sub-m205610', 'sub-m023917',
'sub-m310073', 'sub-m778290', 'sub-m717470', 'sub-m631090', 'sub-m704693', 'sub-m354066', 'sub-m772796', 'sub-m094254', 'sub-m698534', 'sub-m063690', 'sub-m757043', 'sub-m556894',
'sub-m595577', 'sub-m573737', 'sub-m168132', 'sub-m356340', 'sub-m356026', 'sub-m816146', 'sub-m751383', 'sub-m944619', 'sub-m663069', 'sub-m698817', 'sub-m126053', 'sub-m621782',
'sub-m909606', 'sub-m508941', 'sub-m673334', 'sub-m785774', 'sub-m978546', 'sub-m085197', 'sub-m312155', 'sub-m492109', 'sub-m798409', 'sub-m104714', 'sub-m993488', 'sub-m751075',
'sub-m040509', 'sub-m843491', 'sub-m949797', 'sub-m977227', 'sub-m469393', 'sub-m558234', 'sub-m474555', 'sub-m878455', 'sub-m043194', 'sub-m664123', 'sub-m527202', 'sub-m029034',
'sub-m087754', 'sub-m545924', 'sub-m809689', 'sub-m779887', 'sub-m403171', 'sub-m275864', 'sub-m569425', 'sub-m729353', 'sub-m617186', 'sub-m701054', 'sub-m333631', 'sub-m315309',
'sub-m027847', 'sub-m707324', 'sub-m397667', 'sub-m339845', 'sub-m941876', 'sub-m841476', 'sub-m846990', 'sub-m870870', 'sub-m251271', 'sub-m243881', 'sub-m220667', 'sub-m124504',
'sub-m172680', 'sub-m901378', 'sub-m245390', 'sub-m886317', 'sub-m094503', 'sub-m979943', 'sub-m640779', 'sub-m493131', 'sub-m379862', 'sub-m438239', 'sub-m730546', 'sub-m762599',
'sub-m781551', 'sub-m072533', 'sub-m189434', 'sub-m115467', 'sub-m438273', 'sub-m838420', 'sub-m986156', 'sub-m644597', 'sub-m819426', 'sub-m329161', 'sub-m479421', 'sub-m684459',
'sub-m504077', 'sub-m157227', 'sub-m551363', 'sub-m034619', 'sub-m412427', 'sub-m037477', 'sub-m292834']
Min/max sizes of the images and labels
In order to decide on a optimal cropping size, it is useful to know what largest and smallest dimensions across all the subjects. Hence,
min along each dimension: [ 18 42 201]
max along each dimension: [160 175 713]
EDIT: Note that these dimensions correspond to the sizes of the preprocessed images that have been cropped using the spinal cord segmentation mask.