New training with tightly cropped data

[jcohenadad]

spin-off of #65

Stats on processed data

(base) p101317@rosenberg:~/data_nvme_p101317/data_seg_mp2rage_20230210_144231$ python ~/code/model_seg_ms_mp2rage/preprocessing/qc_preprocess.py -s .
Iterating over Subjects: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 283/283 [02:05<00:00,  2.26it/s]
RESOLUTIONS:  Counter({(1.0, 1.0, 1.0): 283})
SIZES:  Counter({(240, 256, 176): 283})
CROP SIZES:  Counter({(20, 88, 21): 3, (19, 89, 24): 2, (30, 94, 22): 2, (24, 98, 23): 2, (21, 77, 19): 2, (23, 86, 21): 2, (26, 98, 20): 2, (19, 91, 20): 2, (24, 101, 20): 2, (21, 93, 22): 2, (19, 94, 20): 2, (24, 83, 24): 2, (21, 97, 21): 2, (20, 96, 21): 2, (20, 104, 23): 2, (26, 95, 21): 2, (21, 102, 20): 2, (18, 99, 21): 2, (20, 78, 21): 1, (24, 108, 24): 1, (26, 102, 19): 1, (18, 98, 20): 1, (27, 105, 21): 1, (26, 94, 19): 1, (24, 92, 20): 1, (25, 107, 21): 1, (18, 108, 19): 1, (27, 105, 20): 1, (28, 78, 20): 1, (25, 105, 22): 1, (20, 81, 22): 1, (22, 93, 22): 1, (25, 88, 21): 1, (24, 77, 22): 1, (19, 81, 25): 1, (22, 93, 19): 1, (18, 96, 19): 1, (29, 94, 22): 1, (23, 104, 21): 1, (23, 106, 22): 1, (22, 85, 20): 1, (21, 88, 21): 1, (25, 92, 21): 1, (31, 100, 19): 1, (28, 91, 24): 1, (30, 98, 19): 1, (25, 97, 26): 1, (22, 102, 22): 1, (23, 100, 20): 1, (17, 83, 21): 1, (21, 92, 19): 1, (18, 90, 21): 1, (28, 103, 20): 1, (21, 86, 21): 1, (23, 89, 21): 1, (29, 94, 20): 1, (19, 93, 21): 1, (23, 97, 20): 1, (26, 108, 23): 1, (21, 90, 22): 1, (23, 78, 20): 1, (31, 107, 22): 1, (37, 100, 22): 1, (34, 88, 20): 1, (19, 86, 20): 1, (27, 88, 20): 1, (32, 102, 23): 1, (23, 75, 20): 1, (37, 86, 23): 1, (24, 91, 18): 1, (19, 91, 22): 1, (18, 92, 25): 1, (24, 95, 22): 1, (17, 93, 21): 1, (31, 89, 23): 1, (20, 93, 22): 1, (19, 88, 25): 1, (22, 90, 20): 1, (33, 102, 21): 1, (23, 91, 20): 1, (18, 94, 20): 1, (20, 96, 22): 1, (18, 79, 20): 1, (21, 85, 23): 1, (19, 106, 23): 1, (17, 101, 18): 1, (30, 97, 19): 1, (19, 104, 24): 1, (35, 125, 21): 1, (23, 76, 19): 1, (20, 86, 28): 1, (21, 101, 19): 1, (19, 83, 22): 1, (23, 93, 19): 1, (16, 83, 20): 1, (18, 90, 19): 1, (21, 83, 20): 1, (18, 98, 22): 1, (23, 93, 20): 1, (26, 86, 20): 1, (31, 94, 23): 1, (30, 115, 20): 1, (20, 91, 21): 1, (27, 96, 22): 1, (25, 91, 22): 1, (18, 93, 19): 1, (22, 96, 21): 1, (26, 83, 23): 1, (38, 101, 24): 1, (30, 94, 23): 1, (29, 108, 18): 1, (29, 113, 21): 1, (24, 84, 25): 1, (25, 92, 23): 1, (26, 98, 19): 1, (29, 89, 24): 1, (24, 101, 18): 1, (19, 98, 18): 1, (24, 73, 22): 1, (34, 94, 24): 1, (24, 84, 19): 1, (26, 97, 22): 1, (19, 105, 27): 1, (24, 99, 21): 1, (28, 99, 24): 1, (21, 100, 20): 1, (28, 83, 23): 1, (34, 88, 21): 1, (21, 88, 19): 1, (21, 78, 22): 1, (19, 95, 20): 1, (17, 96, 19): 1, (19, 93, 19): 1, (20, 101, 21): 1, (19, 69, 18): 1, (20, 84, 22): 1, (32, 95, 20): 1, (25, 101, 24): 1, (26, 86, 21): 1, (34, 83, 24): 1, (29, 75, 19): 1, (22, 84, 20): 1, (41, 85, 21): 1, (30, 91, 18): 1, (19, 89, 20): 1, (26, 89, 19): 1, (19, 89, 25): 1, (21, 99, 26): 1, (24, 99, 18): 1, (27, 65, 20): 1, (19, 93, 22): 1, (22, 88, 19): 1, (21, 89, 20): 1, (27, 91, 21): 1, (26, 96, 22): 1, (25, 116, 19): 1, (31, 97, 24): 1, (27, 97, 21): 1, (24, 94, 21): 1, (26, 88, 20): 1, (22, 95, 23): 1, (22, 99, 18): 1, (22, 102, 24): 1, (39, 98, 21): 1, (25, 103, 20): 1, (37, 104, 24): 1, (24, 74, 20): 1, (30, 96, 26): 1, (27, 99, 23): 1, (23, 102, 20): 1, (30, 80, 22): 1, (18, 102, 20): 1, (32, 106, 21): 1, (20, 97, 22): 1, (21, 107, 20): 1, (24, 97, 22): 1, (23, 92, 18): 1, (31, 99, 24): 1, (29, 102, 21): 1, (18, 87, 31): 1, (29, 109, 22): 1, (21, 99, 18): 1, (29, 83, 25): 1, (19, 84, 19): 1, (25, 109, 22): 1, (26, 69, 27): 1, (22, 70, 20): 1, (27, 106, 19): 1, (20, 96, 23): 1, (36, 93, 18): 1, (22, 95, 20): 1, (53, 91, 26): 1, (26, 92, 20): 1, (24, 92, 24): 1, (25, 102, 20): 1, (24, 99, 23): 1, (16, 76, 25): 1, (19, 90, 21): 1, (25, 99, 22): 1, (23, 101, 20): 1, (30, 74, 20): 1, (18, 89, 21): 1, (22, 85, 21): 1, (25, 89, 25): 1, (22, 103, 22): 1, (20, 88, 20): 1, (30, 92, 23): 1, (27, 106, 20): 1, (20, 95, 22): 1, (26, 91, 21): 1, (31, 86, 21): 1, (22, 101, 19): 1, (17, 87, 22): 1, (17, 100, 21): 1, (19, 90, 23): 1, (25, 88, 19): 1, (23, 97, 21): 1, (21, 100, 22): 1, (21, 95, 20): 1, (39, 75, 18): 1, (21, 99, 22): 1, (22, 81, 20): 1, (18, 87, 19): 1, (19, 89, 22): 1, (19, 107, 20): 1, (21, 98, 24): 1, (22, 85, 19): 1, (22, 85, 24): 1, (18, 99, 22): 1, (23, 96, 24): 1, (28, 99, 21): 1, (28, 95, 22): 1, (25, 96, 20): 1, (22, 96, 25): 1, (33, 93, 21): 1, (22, 94, 21): 1, (25, 93, 21): 1, (20, 91, 19): 1, (17, 101, 19): 1, (27, 91, 24): 1, (24, 96, 21): 1, (25, 95, 21): 1, (24, 106, 20): 1, (30, 91, 24): 1, (23, 84, 21): 1, (18, 88, 22): 1, (28, 106, 19): 1, (20, 97, 23): 1, (18, 96, 20): 1, (18, 75, 23): 1, (21, 87, 17): 1, (27, 105, 25): 1, (31, 97, 21): 1, (25, 91, 19): 1, (26, 86, 23): 1, (27, 81, 20): 1, (22, 96, 26): 1, (30, 104, 20): 1, (18, 94, 21): 1, (19, 87, 20): 1, (23, 84, 25): 1, (24, 85, 25): 1, (22, 92, 24): 1, (19, 96, 20): 1})
List of missing ground truth:  ['sub-C001', 'sub-C002', 'sub-C003', 'sub-C005', 'sub-C006', 'sub-C007', 'sub-C008', 'sub-C009', 'sub-C010', 'sub-C011', 'sub-C012', 'sub-C013', 'sub-C014', 'sub-C015', 'sub-C016', 'sub-C018', 'sub-C019', 'sub-C020', 'sub-C021', 'sub-C022', 'sub-C023', 'sub-C024', 'sub-C025', 'sub-C026', 'sub-C027', 'sub-C028', 'sub-C029', 'sub-C030', 'sub-C031', 'sub-C032', 'sub-C033', 'sub-C034', 'sub-C035', 'sub-C036', 'sub-C037', 'sub-C039', 'sub-C040', 'sub-C041', 'sub-C042', 'sub-C043', 'sub-C044', 'sub-C045', 'sub-C046', 'sub-C047', 'sub-C048', 'sub-C049', 'sub-C050', 'sub-C051', 'sub-C052', 'sub-C053', 'sub-C054', 'sub-C055', 'sub-C056', 'sub-C057', 'sub-C058', 'sub-C059', 'sub-C060', 'sub-C061', 'sub-C062', 'sub-C063', 'sub-C064', 'sub-C065', 'sub-C066', 'sub-C067', 'sub-C068', 'sub-C069', 'sub-C070', 'sub-C071', 'sub-C072', 'sub-C073', 'sub-C074', 'sub-C075', 'sub-C077', 'sub-C078', 'sub-C079', 'sub-C080', 'sub-C081', 'sub-C082', 'sub-C083', 'sub-C084', 'sub-C085', 'sub-C086', 'sub-C087', 'sub-C088', 'sub-C089', 'sub-C090', 'sub-C091', 'sub-C092', 'sub-C093', 'sub-C094', 'sub-C095', 'sub-C096', 'sub-C097', 'sub-C098', 'sub-C099', 'sub-C100', 'sub-C101', 'sub-C102', 'sub-C103', 'sub-C105', 'sub-C106', 'sub-C107', 'sub-C108']
Could not find cropped image for the following subjects:  []
Found shape mismatch in images and GTs for the following subjects:  []
ALERT: Lesion(s) from raters cropped during preprocessing for the following subjects:  ['sub-P027', 'sub-P047', 'sub-P075', 'sub-P130', 'sub-P153', 'sub-P241']

The cropped data size suggest that an appropriate CenterCrop is: 64x128x32, as currently used:

model_seg_ms_mp2rage/config/config_seg_lesion.json

Line 115 in 54d0d00

"CenterCrop": {"size": [32, 64, 128]},

Note
The swapped dimensions might be related to the reorientation of the image in ivadomed.

[jcohenadad]

Interestingly, the tightly cropped data (in green) give lower performance:

[jcohenadad]

Re-processing, this time with larger dilation (back to 32, in gray):

Results make sense with previous mask-dilation based cropping (orange curve).

[jcohenadad]

Closing because this is now fixed-- however there is still an issue about the sensitivity of the cropping window #64