identity dict | complete bool | total_volumes int64 | backend string | gpu_validation dict | splits dict | training_compatibility string | redistribution string |
|---|---|---|---|---|---|---|---|
{
"source_repo": "tqhuyen/harvard-oct-glaucoma-200",
"source_revision": "939a38876b7b9313162842ef2d44b7edc2b57020",
"sources": {
"Training": {
"shape": [
2100,
1,
200,
200,
200
],
"dtype": "uint8",
"offset": 128,
"sample_bytes": 8000000,
... | true | 3,300 | gpu | {
"status": "passed",
"split": "Training",
"index": 276,
"mismatch_bytes": 0
} | {
"Training": {
"identity": {
"source_repo": "tqhuyen/harvard-oct-glaucoma-200",
"source_revision": "939a38876b7b9313162842ef2d44b7edc2b57020",
"sources": {
"Training": {
"shape": [
2100,
1,
200,
200,
200
],
... | Portable export; validate/import into the training cache before use. Local .complete.pt identities are not portable. | Check the upstream dataset terms before publishing. No HF upload was performed. |
Harvard-GF 200^3 — bilateral-filtered volumes
Bilateral-filtered counterpart of
tqhuyen/harvard-oct-glaucoma-200.
Volumes remain at the canonical 200^3 resolution, stored as uint8 .npy.
Provenance
- Raw source:
tqhuyen/harvard-oct-glaucoma-200, revision939a38876b7b9313162842ef2d44b7edc2b57020(itself derived fromharvardairobotics/Harvard-GF, IEEE TMI 2024, Luo et al.). - Filter: grayscale bilateral, applied slice-by-slice along source axis 0.
- Parameters:
sigma_color=0.10,sigma_spatial=4.0,bins=10000,mode=constant,cval=0,win_size=25(auto),channel_axis=None. - Reference implementation:
skimage.restoration.denoise_bilateral(scikit-image 0.26.0). - Quantization:
float32 / 255 -> filter -> clip(0,255) -> uint8(truncation). - Numerical backend: a validated CUDA port produced every output; it matched the scikit-image reference bit-for-bit (0 differing bytes) on a full held-out volume before the run, and only uint8-equal results were written.
- Labels and split membership are copied unchanged from the raw source.
Files
| File | Shape | Dtype |
|---|---|---|
Training_volumes_dn.npy |
(2100, 1, 200, 200, 200) |
uint8 |
Validation_volumes_dn.npy |
(300, 1, 200, 200, 200) |
uint8 |
Test_volumes_dn.npy |
(900, 1, 200, 200, 200) |
uint8 |
{Training,Validation,Test}_labels.npy |
(N,) |
int64 |
Split sizes and positive/negative counts match the raw source exactly
(Training 1083/1017, Validation 176/124, Test 489/411). Downstream
normalization is x / 255.0.
manifest.json records the source revision, filter parameters, implementation
hashes, and the SHA-256 of each volume array and label file.
Intended use and license
Research use only; not for clinical decision-making. This dataset is a derivative of Harvard-GF, which is released under CC BY-NC-ND 4.0 (non-commercial, no derivatives). Treat redistribution accordingly and confirm the upstream terms before making this repository public.
- Downloads last month
- 57