LuGot16 commited on
Commit
93531b0
·
verified ·
1 Parent(s): e0ee1e6

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +91 -28
README.md CHANGED
@@ -65,10 +65,10 @@ Metrics computed on the post-processed predicted masks vs. reference masks; vali
65
  | Item | Value |
66
  |-----------------------|-------------------------------------------------------------------|
67
  | Tissue / species | Mouse testis (*Mus musculus*), CF-1 strain; seminiferous tubules, H&E |
68
- | Dataset | `LuGot16/tubules` (HuggingFace) |
69
- | Train | 273 images — 11 animals, 2 slides per animal |
70
- | Train / Validation | 273 / 49 images, random 85/15 split (seed 42) |
71
- | Split level | Image-level random split (not grouped by animal) |
72
  | Scale | 0.32 µm/px |
73
  | Input | 512×512, ImageNet normalization |
74
  | Classes | 0 = background, 1 = epithelium, 2 = lumen |
@@ -80,31 +80,94 @@ Metrics computed on the post-processed predicted masks vs. reference masks; vali
80
 
81
  ## 4. Robustness, known biases, and scope
82
 
83
- **Generalization (animals / stain intensities / fixative):**
84
- - **Note on validation:** the train/validation split is random at the image level, not
85
- grouped by animal — so images from the same animal may appear in both sets. The validation
86
- metrics in Sections 1–2 may therefore be optimistic with respect to generalization to fully
87
- unseen animals. The strongest evidence for cross-animal generalization is the
88
- external test below (129 images from other animals and experiments, with different stains
89
- and fixatives).
90
- - **External test**: 129 images from **other animals and different experiments** (the training
91
- set comprises 11 animals across 3 experiments; the external images come from outside that set),
92
- with widely varying stain intensities and different fixatives. Satisfactory segmentation on
93
- **128/129**; the single failure was a patch of extremely thin epithelium left unsegmented in
94
- one image. Macenko stain normalization during training contributes to this tolerance to stain
95
- and fixation variation.
96
-
97
- **Known biases:**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
98
  - Boundary accuracy vs. reference masks (validation, n = 49):
99
- - Tubule outer boundary: ASSD 0.59 µm, HD95 1.54 µm, mean signed displacement +0.13 µm (Boundary-F1 @1.6 µm = 0.97).
100
- - Lumen boundary: ASSD 1.47 µm, HD95 4.99 µm, mean signed displacement −0.05 µm (Boundary-F1 @1.6 µm = 0.75); the larger local variation reflects the convoluted scalloped lumen contour, with no systematic offset.
101
-
102
- The model reproduces the reference masks with sub-pixel mean boundary agreement (signed displacement ≈ 0 for both structures; a +0.13 µm shift accounts for ~0.25% area). The ~+12% / +8% area difference vs. manual ImageJ freehand tracing is therefore a convention difference between the reference masks and the freehand protocol — faithfully inherited by the model — not error introduced by the model. It is systematic (r ≈ 0.99) and preserves shape (AR/roundness CCC ≈ 0.99).
103
- - **Lumen perimeter**: good mean agreement (CCC 0.93, bias ≈ 0) but wide LoA (±110 µm):
104
- varies case by case due to tracing convention.
105
-
106
- **Validated scope:** single-tubule crops at 0.32 µm/px, H&E staining. **Not validated** on
107
- multi-tubule fields or other magnifications.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
108
 
109
  ---
110
 
 
65
  | Item | Value |
66
  |-----------------------|-------------------------------------------------------------------|
67
  | Tissue / species | Mouse testis (*Mus musculus*), CF-1 strain; seminiferous tubules, H&E |
68
+ | Dataset | `LuGot16/tubules` (HuggingFace) — 322 images |
69
+ | Source material | 11 animals × 2 slides per animal, across 3 experiments |
70
+ | Train / Validation | 273 / 49 images, random 85/15 split (seed 42) |
71
+ | Split level | Image-level random split (not grouped by animal) |
72
  | Scale | 0.32 µm/px |
73
  | Input | 512×512, ImageNet normalization |
74
  | Classes | 0 = background, 1 = epithelium, 2 = lumen |
 
80
 
81
  ## 4. Robustness, known biases, and scope
82
 
83
+ ### Generalization — external validation (other animals, stains, fixatives)
84
+
85
+ The train/validation split is random at the image level, not grouped by animal, so images
86
+ from the same animal may appear in both sets; the validation metrics in Sections 1–2 may be
87
+ optimistic for fully unseen animals. The primary evidence for cross-animal generalization is
88
+ an external set of **129 images from other animals and experiments**, with different stains and
89
+ fixatives (~0.32 µm/px). On a hand-traced subset, the model was evaluated quantitatively.
90
+
91
+ **Segmentation (per-image mean, n = 29):**
92
+
93
+ | Class | IoU | Dice | clDice |
94
+ |----------------|-------|-------|--------|
95
+ | Background | 0.986 | 0.993 | — |
96
+ | Epithelium | 0.885 | 0.938 | 0.948 |
97
+ | Lumen | 0.867 | 0.927 | 0.927 |
98
+ | **Mean** | **mIoU 0.913** | **0.953** | — |
99
+
100
+ Per-image mIoU 0.913 (median 0.924, min 0.811) vs. 0.953 in-domain — a modest, expected
101
+ cross-domain drop.
102
+
103
+ **Boundary accuracy (n = 29):**
104
+
105
+ | Boundary (µm) | ASSD | HD95 | signed disp. | BF @1.6 µm |
106
+ |---------------|------|-------|--------------|------------|
107
+ | Tubule | 0.97 | 3.06 | +0.48 | 0.85 |
108
+ | Lumen | 3.86 | 13.90 | +0.72 | 0.53 |
109
+
110
+ **Morphometric agreement vs. manual ImageJ (n = 29):**
111
+
112
+ *Tubule*
113
+
114
+ | Descriptor | Pearson r | CCC | Rel. error | Scale factor |
115
+ |------------|-----------|-----|------------|--------------|
116
+ | Area | 0.998 | 0.769 | 14.3% | 1.142 |
117
+ | Perimeter | 0.966 | 0.819 | 5.6% | 1.061 |
118
+ | Major axis | 0.997 | 0.852 | 6.9% | 1.069 |
119
+ | Minor axis | 0.997 | 0.753 | 7.4% | 1.072 |
120
+ | Max Feret | 0.996 | 0.863 | 6.3% | 1.064 |
121
+ | Min Feret | 0.993 | 0.756 | 7.1% | 1.069 |
122
+ | Aspect ratio | 0.995 | 0.994 | 0.8% | 0.997 |
123
+ | Roundness | 0.993 | 0.993 | 0.9% | 0.999 |
124
+
125
+ *Lumen*
126
+
127
+ | Descriptor | Pearson r | CCC | Rel. error | Scale factor |
128
+ |------------|-----------|-----|------------|--------------|
129
+ | Area | 0.955 | 0.892 | 14.0% | 1.110 |
130
+ | Perimeter | 0.790 | 0.681 | 13.4% | 0.909 |
131
+ | Major axis | 0.941 | 0.842 | 9.1% | 1.077 |
132
+ | Minor axis | 0.939 | 0.759 | 11.1% | 1.101 |
133
+ | Max Feret | 0.924 | 0.917 | 5.1% | 1.010 |
134
+ | Min Feret | 0.935 | 0.915 | 5.3% | 1.034 |
135
+ | Aspect ratio | 0.925 | 0.881 | 4.1% | 0.983 |
136
+ | Roundness | 0.893 | 0.802 | 6.0% | 0.954 |
137
+
138
+ The size bias (+6–14%) and the near-perfect agreement of the shape descriptors
139
+ (tubule AR/roundness CCC ≈ 0.99, scale factor ≈ 1.00) reproduce the in-domain pattern,
140
+ confirming the offset is a **domain-stable** scaling difference, not out-of-domain distortion.
141
+ Lumen perimeter is the only noisy metric (r 0.79), reflecting the scalloped, low-contrast
142
+ lumen boundary.
143
+
144
+ ### Known biases
145
+
146
  - Boundary accuracy vs. reference masks (validation, n = 49):
147
+ - Tubule: ASSD 0.59 µm, HD95 1.54 µm, mean signed displacement +0.13 µm (Boundary-F1 @1.6 µm = 0.97).
148
+ - Lumen: ASSD 1.47 µm, HD95 4.99 µm, mean signed displacement −0.05 µm (Boundary-F1 @1.6 µm = 0.75).
149
+
150
+ The model reproduces the reference masks with sub-pixel mean boundary agreement; the
151
+ external boundary displacement stays small (tubule +0.48 µm ≈ 1.5 px). The ~+12% / +8%
152
+ (tubule / lumen) area difference vs. manual ImageJ **freehand** tracing is therefore a
153
+ convention difference between the reference masks and the freehand protocol — faithfully
154
+ inherited by the model — not error introduced by it. It is systematic (r ≈ 0.99) and
155
+ preserves shape (**tubule** AR/roundness CCC ≈ 0.99; lumen roundness lower, CCC 0.82).
156
+
157
+ - **Lumen perimeter**: good mean agreement in-domain (CCC 0.93, bias ≈ 0) but wide LoA
158
+ (±110 µm), varying case by case with tracing convention.
159
+
160
+ ### Failure mode
161
+
162
+ One badly-fixed external tubule had its lumen left unsegmented. This was **correctly flagged by
163
+ the model's confidence/status output** (and is a clear outlier — boundary displacement −10 µm),
164
+ so it was excluded from the agreement statistics. Gross fixation artifacts are surfaced by the
165
+ confidence output rather than silently mis-segmented.
166
+
167
+ ### Validated scope
168
+
169
+ Single-tubule crops at 0.32 µm/px, H&E staining. **Not validated** on multi-tubule fields or
170
+ other magnifications.
171
 
172
  ---
173