shawn24 commited on
Commit
8388d8b
·
verified ·
1 Parent(s): 2b54c62

Add portrait 15-epoch classification checkpoints

Browse files

Add RSNA and VinDr density, CMMD cancer/mass/calcification, and RSNA conditional invasive-cancer inference checkpoints with integrity and input metadata.

Files changed (36) hide show
  1. fullfinetune_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/README.md +9 -0
  2. fullfinetune_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_calcification_full_finetune_seed10_fold0_best_validation_auroc.pth +3 -0
  3. fullfinetune_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/release_metadata.json +33 -0
  4. fullfinetune_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/README.md +9 -0
  5. fullfinetune_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth +3 -0
  6. fullfinetune_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/release_metadata.json +33 -0
  7. fullfinetune_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/README.md +9 -0
  8. fullfinetune_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_mass_full_finetune_seed10_fold0_best_validation_auroc.pth +3 -0
  9. fullfinetune_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/release_metadata.json +33 -0
  10. fullfinetune_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/README.md +9 -0
  11. fullfinetune_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_density_full_finetune_seed10_fold0_best_validation_auroc.pth +3 -0
  12. fullfinetune_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/release_metadata.json +35 -0
  13. fullfinetune_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/README.md +9 -0
  14. fullfinetune_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_invasive_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth +3 -0
  15. fullfinetune_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/release_metadata.json +35 -0
  16. fullfinetune_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/README.md +9 -0
  17. fullfinetune_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/checkpoint/mammo_fm_vindr_density_full_finetune_seed10_fold0_best_validation_auroc.pth +3 -0
  18. fullfinetune_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/release_metadata.json +35 -0
  19. linear_probe_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/README.md +9 -0
  20. linear_probe_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_calcification_linear_probe_seed10_fold0_best_validation_auroc.pth +3 -0
  21. linear_probe_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/release_metadata.json +33 -0
  22. linear_probe_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/README.md +9 -0
  23. linear_probe_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth +3 -0
  24. linear_probe_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/release_metadata.json +33 -0
  25. linear_probe_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/README.md +9 -0
  26. linear_probe_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_mass_linear_probe_seed10_fold0_best_validation_auroc.pth +3 -0
  27. linear_probe_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/release_metadata.json +33 -0
  28. linear_probe_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/README.md +9 -0
  29. linear_probe_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_density_linear_probe_seed10_fold0_best_validation_auroc.pth +3 -0
  30. linear_probe_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/release_metadata.json +35 -0
  31. linear_probe_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/README.md +9 -0
  32. linear_probe_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_invasive_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth +3 -0
  33. linear_probe_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/release_metadata.json +35 -0
  34. linear_probe_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/README.md +9 -0
  35. linear_probe_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/checkpoint/mammo_fm_vindr_density_linear_probe_seed10_fold0_best_validation_auroc.pth +3 -0
  36. linear_probe_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/release_metadata.json +35 -0
fullfinetune_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # CMMD calcification — Full fine-tuning
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score.
6
+
7
+ Checkpoint: [`mammo_fm_cmmd_calcification_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_cmmd_calcification_full_finetune_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
fullfinetune_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_calcification_full_finetune_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c8990edfd0534183867e067826c014075b0c92d057c4e4969e70e043e0b31ea
3
+ size 114481160
fullfinetune_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "CMMD",
4
+ "task": "calcification",
5
+ "method": "full_finetune",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "negative",
9
+ "positive"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 5,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "positive_class": "calcification",
29
+ "checkpoint_file": "checkpoint/mammo_fm_cmmd_calcification_full_finetune_seed10_fold0_best_validation_auroc.pth",
30
+ "checkpoint_sha256": "7c8990edfd0534183867e067826c014075b0c92d057c4e4969e70e043e0b31ea",
31
+ "checkpoint_bytes": 114481160,
32
+ "model_tensor_count": 854
33
+ }
fullfinetune_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # CMMD cancer — Full fine-tuning
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score.
6
+
7
+ Checkpoint: [`mammo_fm_cmmd_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_cmmd_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
fullfinetune_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2c93326d4037cf7a7e30aafe608a69298fecbb7bf7c080d4118033174ee84575
3
+ size 114432338
fullfinetune_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "CMMD",
4
+ "task": "cancer",
5
+ "method": "full_finetune",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "negative",
9
+ "positive"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 13,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "positive_class": "cancer",
29
+ "checkpoint_file": "checkpoint/mammo_fm_cmmd_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth",
30
+ "checkpoint_sha256": "2c93326d4037cf7a7e30aafe608a69298fecbb7bf7c080d4118033174ee84575",
31
+ "checkpoint_bytes": 114432338,
32
+ "model_tensor_count": 854
33
+ }
fullfinetune_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # CMMD mass — Full fine-tuning
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score.
6
+
7
+ Checkpoint: [`mammo_fm_cmmd_mass_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_cmmd_mass_full_finetune_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
fullfinetune_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_mass_full_finetune_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c4a11f1666b3364f1562369b0624f2377e89966c62e717d062d074ce990d9703
3
+ size 114430622
fullfinetune_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "CMMD",
4
+ "task": "mass",
5
+ "method": "full_finetune",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "negative",
9
+ "positive"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 8,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "positive_class": "mass",
29
+ "checkpoint_file": "checkpoint/mammo_fm_cmmd_mass_full_finetune_seed10_fold0_best_validation_auroc.pth",
30
+ "checkpoint_sha256": "c4a11f1666b3364f1562369b0624f2377e89966c62e717d062d074ce990d9703",
31
+ "checkpoint_bytes": 114430622,
32
+ "model_tensor_count": 854
33
+ }
fullfinetune_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # RSNA density — Full fine-tuning
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply softmax to the four logits to obtain density probabilities in A/B/C/D order.
6
+
7
+ Checkpoint: [`mammo_fm_rsna_density_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_rsna_density_full_finetune_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
fullfinetune_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_density_full_finetune_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9c2083f67a6cad82f04460048659f6f2004360946f08f8401ec906a9a7014cb7
3
+ size 114496492
fullfinetune_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "RSNA",
4
+ "task": "density",
5
+ "method": "full_finetune",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "A",
9
+ "B",
10
+ "C",
11
+ "D"
12
+ ],
13
+ "num_classes": 4,
14
+ "feature_dimension": 2048,
15
+ "input_channels": 3,
16
+ "image_size_hw": [
17
+ 1520,
18
+ 912
19
+ ],
20
+ "output_activation": "softmax",
21
+ "training_epochs": 15,
22
+ "selected_epoch": 4,
23
+ "checkpoint_selection": "best validation macro one-vs-rest AUROC",
24
+ "seed": 10,
25
+ "fold": 0,
26
+ "data_fraction": 1.0,
27
+ "normalization_mean": 0.3089279,
28
+ "normalization_std": 0.25053555408335154,
29
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
30
+ "score_semantics": "four mutually exclusive density class probabilities in A, B, C, D order",
31
+ "checkpoint_file": "checkpoint/mammo_fm_rsna_density_full_finetune_seed10_fold0_best_validation_auroc.pth",
32
+ "checkpoint_sha256": "9c2083f67a6cad82f04460048659f6f2004360946f08f8401ec906a9a7014cb7",
33
+ "checkpoint_bytes": 114496492,
34
+ "model_tensor_count": 854
35
+ }
fullfinetune_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # RSNA invasive cancer — Full fine-tuning
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score. This classifier distinguishes invasive from non-invasive cancer among cancer-positive cases only. It is not a general cancer detector or an estimate of population invasive cancer risk.
6
+
7
+ Checkpoint: [`mammo_fm_rsna_invasive_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_rsna_invasive_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
fullfinetune_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_invasive_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b48f0f4c0e8f6768aa29426396baad8c36cf5cb27be768c4b19949a28732c2f0
3
+ size 114483068
fullfinetune_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "RSNA",
4
+ "task": "invasive_cancer",
5
+ "method": "full_finetune",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "non_invasive_cancer",
9
+ "invasive_cancer"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 6,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "intended_cohort": "cancer-positive cases only",
29
+ "positive_class": "invasive_cancer",
30
+ "score_semantics": "conditional invasive-versus-non-invasive cancer score; not general-population invasive cancer risk",
31
+ "checkpoint_file": "checkpoint/mammo_fm_rsna_invasive_cancer_full_finetune_seed10_fold0_best_validation_auroc.pth",
32
+ "checkpoint_sha256": "b48f0f4c0e8f6768aa29426396baad8c36cf5cb27be768c4b19949a28732c2f0",
33
+ "checkpoint_bytes": 114483068,
34
+ "model_tensor_count": 854
35
+ }
fullfinetune_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # VINDR density — Full fine-tuning
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply softmax to the four logits to obtain density probabilities in A/B/C/D order.
6
+
7
+ Checkpoint: [`mammo_fm_vindr_density_full_finetune_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_vindr_density_full_finetune_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
fullfinetune_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/checkpoint/mammo_fm_vindr_density_full_finetune_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:419f0294dcf80bc6a802ea4ba877bcacbe7a0ebcf66ea09560f3a0d5a5b364df
3
+ size 114501062
fullfinetune_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "VINDR",
4
+ "task": "density",
5
+ "method": "full_finetune",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "A",
9
+ "B",
10
+ "C",
11
+ "D"
12
+ ],
13
+ "num_classes": 4,
14
+ "feature_dimension": 2048,
15
+ "input_channels": 3,
16
+ "image_size_hw": [
17
+ 1520,
18
+ 912
19
+ ],
20
+ "output_activation": "softmax",
21
+ "training_epochs": 15,
22
+ "selected_epoch": 2,
23
+ "checkpoint_selection": "best validation macro one-vs-rest AUROC",
24
+ "seed": 10,
25
+ "fold": 0,
26
+ "data_fraction": 1.0,
27
+ "normalization_mean": 0.3089279,
28
+ "normalization_std": 0.25053555408335154,
29
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
30
+ "score_semantics": "four mutually exclusive density class probabilities in A, B, C, D order",
31
+ "checkpoint_file": "checkpoint/mammo_fm_vindr_density_full_finetune_seed10_fold0_best_validation_auroc.pth",
32
+ "checkpoint_sha256": "419f0294dcf80bc6a802ea4ba877bcacbe7a0ebcf66ea09560f3a0d5a5b364df",
33
+ "checkpoint_bytes": 114501062,
34
+ "model_tensor_count": 854
35
+ }
linear_probe_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # CMMD calcification — Linear probe
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score.
6
+
7
+ Checkpoint: [`mammo_fm_cmmd_calcification_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_cmmd_calcification_linear_probe_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
linear_probe_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_calcification_linear_probe_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:45ba7ad249e12e3348252acae5325c44abf63b9e37f07ff4e8b2772f47428010
3
+ size 114480302
linear_probe_mammo_FM_image_encoder/cmmd/calcification_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "CMMD",
4
+ "task": "calcification",
5
+ "method": "linear_probe",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "negative",
9
+ "positive"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 15,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "positive_class": "calcification",
29
+ "checkpoint_file": "checkpoint/mammo_fm_cmmd_calcification_linear_probe_seed10_fold0_best_validation_auroc.pth",
30
+ "checkpoint_sha256": "45ba7ad249e12e3348252acae5325c44abf63b9e37f07ff4e8b2772f47428010",
31
+ "checkpoint_bytes": 114480302,
32
+ "model_tensor_count": 854
33
+ }
linear_probe_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # CMMD cancer — Linear probe
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score.
6
+
7
+ Checkpoint: [`mammo_fm_cmmd_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_cmmd_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
linear_probe_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b2d73669ac003a68332bff328d8f321da017f02ee84057370ae3f4a6db47749e
3
+ size 114431480
linear_probe_mammo_FM_image_encoder/cmmd/cancer_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "CMMD",
4
+ "task": "cancer",
5
+ "method": "linear_probe",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "negative",
9
+ "positive"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 14,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "positive_class": "cancer",
29
+ "checkpoint_file": "checkpoint/mammo_fm_cmmd_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth",
30
+ "checkpoint_sha256": "b2d73669ac003a68332bff328d8f321da017f02ee84057370ae3f4a6db47749e",
31
+ "checkpoint_bytes": 114431480,
32
+ "model_tensor_count": 854
33
+ }
linear_probe_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # CMMD mass — Linear probe
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score.
6
+
7
+ Checkpoint: [`mammo_fm_cmmd_mass_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_cmmd_mass_linear_probe_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
linear_probe_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/checkpoint/mammo_fm_cmmd_mass_linear_probe_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a2941e1c95b02dd91f5cb04ba81132aca8e70d51fdb29c70d547665a6c05cb83
3
+ size 114429700
linear_probe_mammo_FM_image_encoder/cmmd/mass_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "CMMD",
4
+ "task": "mass",
5
+ "method": "linear_probe",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "negative",
9
+ "positive"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 15,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "positive_class": "mass",
29
+ "checkpoint_file": "checkpoint/mammo_fm_cmmd_mass_linear_probe_seed10_fold0_best_validation_auroc.pth",
30
+ "checkpoint_sha256": "a2941e1c95b02dd91f5cb04ba81132aca8e70d51fdb29c70d547665a6c05cb83",
31
+ "checkpoint_bytes": 114429700,
32
+ "model_tensor_count": 854
33
+ }
linear_probe_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # RSNA density — Linear probe
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply softmax to the four logits to obtain density probabilities in A/B/C/D order.
6
+
7
+ Checkpoint: [`mammo_fm_rsna_density_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_rsna_density_linear_probe_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
linear_probe_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_density_linear_probe_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:78496b9e1be093b78665d8841c2be89733effdc6f373db7049578b59f0ad4109
3
+ size 114457042
linear_probe_mammo_FM_image_encoder/rsna/density_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "RSNA",
4
+ "task": "density",
5
+ "method": "linear_probe",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "A",
9
+ "B",
10
+ "C",
11
+ "D"
12
+ ],
13
+ "num_classes": 4,
14
+ "feature_dimension": 2048,
15
+ "input_channels": 3,
16
+ "image_size_hw": [
17
+ 1520,
18
+ 912
19
+ ],
20
+ "output_activation": "softmax",
21
+ "training_epochs": 15,
22
+ "selected_epoch": 15,
23
+ "checkpoint_selection": "best validation macro one-vs-rest AUROC",
24
+ "seed": 10,
25
+ "fold": 0,
26
+ "data_fraction": 1.0,
27
+ "normalization_mean": 0.3089279,
28
+ "normalization_std": 0.25053555408335154,
29
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
30
+ "score_semantics": "four mutually exclusive density class probabilities in A, B, C, D order",
31
+ "checkpoint_file": "checkpoint/mammo_fm_rsna_density_linear_probe_seed10_fold0_best_validation_auroc.pth",
32
+ "checkpoint_sha256": "78496b9e1be093b78665d8841c2be89733effdc6f373db7049578b59f0ad4109",
33
+ "checkpoint_bytes": 114457042,
34
+ "model_tensor_count": 854
35
+ }
linear_probe_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # RSNA invasive cancer — Linear probe
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply sigmoid to the single logit to obtain the binary classification score. This classifier distinguishes invasive from non-invasive cancer among cancer-positive cases only. It is not a general cancer detector or an estimate of population invasive cancer risk.
6
+
7
+ Checkpoint: [`mammo_fm_rsna_invasive_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_rsna_invasive_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
linear_probe_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/checkpoint/mammo_fm_rsna_invasive_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8c0e3ffa0994a9c3baa0c7a8e506b5f8fb0446da65c9e2bd727edc690661f9e
3
+ size 114482210
linear_probe_mammo_FM_image_encoder/rsna/invasive_cancer_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "RSNA",
4
+ "task": "invasive_cancer",
5
+ "method": "linear_probe",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "non_invasive_cancer",
9
+ "invasive_cancer"
10
+ ],
11
+ "num_classes": 1,
12
+ "feature_dimension": 2048,
13
+ "input_channels": 3,
14
+ "image_size_hw": [
15
+ 1520,
16
+ 912
17
+ ],
18
+ "output_activation": "sigmoid",
19
+ "training_epochs": 15,
20
+ "selected_epoch": 14,
21
+ "checkpoint_selection": "best validation AUROC",
22
+ "seed": 10,
23
+ "fold": 0,
24
+ "data_fraction": 1.0,
25
+ "normalization_mean": 0.3089279,
26
+ "normalization_std": 0.25053555408335154,
27
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
28
+ "intended_cohort": "cancer-positive cases only",
29
+ "positive_class": "invasive_cancer",
30
+ "score_semantics": "conditional invasive-versus-non-invasive cancer score; not general-population invasive cancer risk",
31
+ "checkpoint_file": "checkpoint/mammo_fm_rsna_invasive_cancer_linear_probe_seed10_fold0_best_validation_auroc.pth",
32
+ "checkpoint_sha256": "e8c0e3ffa0994a9c3baa0c7a8e506b5f8fb0446da65c9e2bd727edc690661f9e",
33
+ "checkpoint_bytes": 114482210,
34
+ "model_tensor_count": 854
35
+ }
linear_probe_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # VINDR density — Linear probe
2
+
3
+ Mammo-FM EfficientNet-B5 inference checkpoint from a completed 15-epoch portrait run. The model was selected by validation AUROC, using the full training split, seed 10 and fold 0. Input geometry is 1520 pixels high × 912 pixels wide with three channels.
4
+
5
+ Apply softmax to the four logits to obtain density probabilities in A/B/C/D order.
6
+
7
+ Checkpoint: [`mammo_fm_vindr_density_linear_probe_seed10_fold0_best_validation_auroc.pth`](checkpoint/mammo_fm_vindr_density_linear_probe_seed10_fold0_best_validation_auroc.pth). See [`release_metadata.json`](release_metadata.json) for class order, normalization and integrity checks.
8
+
9
+ The checkpoint contains only model tensors and simple inference metadata; load with `torch.load(..., weights_only=True)`. For research use; not a standalone clinical device.
linear_probe_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/checkpoint/mammo_fm_vindr_density_linear_probe_seed10_fold0_best_validation_auroc.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ccd794b0059dbb840b973f5f0dabd42441ed87c79fabd50baffc918d207d76c5
3
+ size 114496492
linear_probe_mammo_FM_image_encoder/vindr/density_classification/portrait_15epochs/release_metadata.json ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "release_schema_version": 1,
3
+ "dataset": "VINDR",
4
+ "task": "density",
5
+ "method": "linear_probe",
6
+ "model": "Mammo-FM EfficientNet-B5 image encoder with linear classification head",
7
+ "class_names": [
8
+ "A",
9
+ "B",
10
+ "C",
11
+ "D"
12
+ ],
13
+ "num_classes": 4,
14
+ "feature_dimension": 2048,
15
+ "input_channels": 3,
16
+ "image_size_hw": [
17
+ 1520,
18
+ 912
19
+ ],
20
+ "output_activation": "softmax",
21
+ "training_epochs": 15,
22
+ "selected_epoch": 15,
23
+ "checkpoint_selection": "best validation macro one-vs-rest AUROC",
24
+ "seed": 10,
25
+ "fold": 0,
26
+ "data_fraction": 1.0,
27
+ "normalization_mean": 0.3089279,
28
+ "normalization_std": 0.25053555408335154,
29
+ "initialization_checkpoint_filename": "Mammo-FM_BatmanlabTrained_CLIP.tar",
30
+ "score_semantics": "four mutually exclusive density class probabilities in A, B, C, D order",
31
+ "checkpoint_file": "checkpoint/mammo_fm_vindr_density_linear_probe_seed10_fold0_best_validation_auroc.pth",
32
+ "checkpoint_sha256": "ccd794b0059dbb840b973f5f0dabd42441ed87c79fabd50baffc918d207d76c5",
33
+ "checkpoint_bytes": 114496492,
34
+ "model_tensor_count": 854
35
+ }