File size: 9,522 Bytes
6d63dbf
 
 
 
 
915f7d5
 
 
 
 
 
 
 
 
 
 
 
 
 
0c1f67f
 
 
 
 
 
915f7d5
0c1f67f
 
 
915f7d5
 
916957e
addf3fb
915f7d5
 
e0ee1e6
915f7d5
 
e0ee1e6
 
 
 
 
 
 
 
 
 
915f7d5
 
e0ee1e6
 
 
 
 
 
 
 
 
 
915f7d5
 
 
 
 
 
 
93531b0
 
 
 
915f7d5
 
 
 
 
 
 
 
 
 
 
93531b0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0c1f67f
93531b0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f2dac56
 
 
 
 
93531b0
 
 
 
 
915f7d5
 
 
 
 
 
 
 
f7712b7
7521291
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
---
license: apache-2.0
pipeline_tag: image-segmentation
---


# TubuleSegmentation v1.0.0 — Seminiferous tubule segmentation and morphometry (mouse testis, H&E)

Semantic segmentation model for seminiferous tubules in H&E sections of mouse testis
(*Mus musculus*, CF-1 strain). Segments 3 classes (0 = background, 1 = epithelium,
2 = lumen) and derives calibrated morphometric metrics. Architecture: EfficientNet-B4
encoder + UNet with a dual decoder (segmentation + boundary) and SCSE attention.
Input 512×512, ImageNet normalization, scale 0.32 µm/px.


---

## 1. Segmentation performance (validation, n = 49)

| Class          | IoU   | Dice / F1 | clDice |
|----------------|-------|-----------|--------|
| Background (0) | 0.990 | 0.995     | —      |
| Epithelium (1) | 0.938 | 0.968     | 0.984  |
| Lumen (2)      | 0.935 | 0.967     | 0.968  |
| Mean           | mIoU 0.954 | 0.977 | —    |

Per-image (n = 49): mIoU 0.953 ± 0.013 (min 0.900, median 0.955);
mDice 0.976 ± 0.007 (min 0.946). No image falls below mIoU 0.90.
Metrics computed on the post-processed predicted masks vs. reference masks; validation set, never train.


![Example segmentation output](https://huggingface.co/LuGot16/Tubules_Segmentation/resolve/main/assets/figure_segmentation_examples_hf.png)

---

## 2. Agreement with manual measurement (ImageJ, gold standard) — validation, n = 48

### Tubule
| Metric | Pearson r | CCC (Lin) [95% CI] | Bias | Rel. error (%) | 95% LoA | MAE | Scale factor [95% CI] |
|---|---|---|---|---|---|---|---|
| Area (µm²) | 0.996 | 0.773 [0.701, 0.818] | +4074.6 | 12.2% | [+2530.3, +5618.9] | 4074.6 | 1.126 [1.108, 1.132] |
| Major axis (µm) | 0.996 | 0.831 [0.775, 0.863] | +14.34 | 6.3% | [+9.17, +19.52] | 14.34 | 1.065 [1.058, 1.066] |
| Minor axis (µm) | 0.995 | 0.782 [0.691, 0.838] | +11.91 | 6.3% | [+8.39, +15.44] | 11.91 | 1.066 [1.058, 1.068] |
| Max Feret (µm) | 0.997 | 0.850 [0.794, 0.882] | +13.43 | 5.9% | [+8.54, +18.32] | 13.43 | 1.058 [1.055, 1.063] |
| Min Feret (µm) | 0.995 | 0.796 [0.702, 0.850] | +11.38 | 6.1% | [+7.62, +15.15] | 11.38 | 1.062 [1.055, 1.066] |
| Perimeter (µm) | 0.996 | 0.790 [0.721, 0.833] | +38.26 | 5.8% | [+26.87, +49.64] | 38.26 | 1.059 [1.052, 1.063] |
| Aspect ratio | 0.999 | 0.999 [0.997, 0.999] | −0.0002 | 0.4% | [−0.013, +0.012] | 0.005 | 0.999 [0.997, 1.002] |
| Roundness | 0.996 | 0.993 [0.988, 0.996] | −0.006 | 0.8% | [−0.020, +0.008] | 0.007 | 0.993 [0.992, 0.995] |

### Lumen
| Metric | Pearson r | CCC (Lin) [95% CI] | Bias | Rel. error (%) | 95% LoA | MAE | Scale factor [95% CI] |
|---|---|---|---|---|---|---|---|
| Area (µm²) | 0.993 | 0.948 [0.922, 0.963] | +1114.7 | 8.1% | [+37.7, +2191.8] | 1114.7 | 1.075 [1.067, 1.091] |
| Major axis (µm) | 0.970 | 0.826 [0.746, 0.877] | +11.87 | 8.0% | [+1.84, +21.90] | 11.87 | 1.074 [1.070, 1.081] |
| Minor axis (µm) | 0.989 | 0.880 [0.800, 0.918] | +9.78 | 8.4% | [+3.69, +15.87] | 9.78 | 1.082 [1.072, 1.098] |
| Max Feret (µm) | 0.966 | 0.950 [0.925, 0.965] | +4.07 | 3.7% | [−7.67, +15.82] | 6.41 | 1.033 [1.027, 1.037] |
| Min Feret (µm) | 0.983 | 0.972 [0.953, 0.982] | +3.06 | 3.1% | [−4.41, +10.53] | 4.16 | 1.031 [1.017, 1.034] |
| Perimeter (µm) | 0.934 | 0.931 [0.889, 0.957] | −6.96 | 5.9% | [−115.5, +101.5] | 41.45 | 1.008 [0.984, 1.028] |
| Aspect ratio | 0.974 | 0.971 [0.933, 0.987] | −0.008 | 2.7% | [−0.108, +0.093] | 0.036 | 0.991 [0.982, 1.004] |
| Roundness | 0.939 | 0.818 [0.709, 0.887] | −0.059 | 7.5% | [−0.138, +0.020] | 0.059 | 0.937 [0.921, 0.948] |
---

## 3. Data and training

| Item                  | Value                                                              |
|-----------------------|-------------------------------------------------------------------|
| Tissue / species      | Mouse testis (*Mus musculus*), CF-1 strain; seminiferous tubules, H&E |
| Dataset               | `LuGot16/tubules` (HuggingFace) — 322 images                       |
| Source material       | 11 animals × 2 slides per animal, across 3 experiments            |
| Train / Validation    | 273 / 49 images, random 85/15 split (seed 42)                     |
| Split level           | Image-level random split (not grouped by animal)                  |
| Scale                 | 0.32 µm/px                                                         |
| Input                 | 512×512, ImageNet normalization                                   |
| Classes               | 0 = background, 1 = epithelium, 2 = lumen                         |
| Augmentation          | Macenko (stain normalization) + geometric                         |
| Post-processing       | largest connected component → morphological closing → hole filling → lumen cleanup |
| Inference             | 8× TTA (4 rotations × 2 flips)                                    |

---

## 4. Robustness, known biases, and scope

### Generalization — external validation (other animals, stains, fixatives)

The train/validation split is random at the image level, not grouped by animal, so images
from the same animal may appear in both sets; the validation metrics in Sections 1–2 may be
optimistic for fully unseen animals. The primary evidence for cross-animal generalization is
an external set of **129 images from other animals and experiments**, with different stains and
fixatives (~0.32 µm/px). On a hand-traced subset, the model was evaluated quantitatively.

**Segmentation (per-image mean, n = 29):**

| Class          | IoU   | Dice  | clDice |
|----------------|-------|-------|--------|
| Background     | 0.986 | 0.993 | —      |
| Epithelium     | 0.885 | 0.938 | 0.948  |
| Lumen          | 0.867 | 0.927 | 0.927  |
| **Mean**       | **mIoU 0.913** | **0.953** | — |

Per-image mIoU 0.913 (median 0.924, min 0.811) vs. 0.953 in-domain — a modest, expected
cross-domain drop.

**Boundary accuracy (n = 29):**

| Boundary (µm) | ASSD | HD95  | signed disp. | BF @1.6 µm |
|---------------|------|-------|--------------|------------|
| Tubule        | 0.97 | 3.06  | +0.48        | 0.85       |
| Lumen         | 3.86 | 13.90 | +0.72        | 0.53       |

**Morphometric agreement vs. manual ImageJ (n = 29):**

*Tubule*

| Descriptor | Pearson r | CCC | Rel. error | Scale factor |
|------------|-----------|-----|------------|--------------|
| Area       | 0.998 | 0.769 | 14.3% | 1.142 |
| Perimeter  | 0.966 | 0.819 | 5.6%  | 1.061 |
| Major axis | 0.997 | 0.852 | 6.9%  | 1.069 |
| Minor axis | 0.997 | 0.753 | 7.4%  | 1.072 |
| Max Feret  | 0.996 | 0.863 | 6.3%  | 1.064 |
| Min Feret  | 0.993 | 0.756 | 7.1%  | 1.069 |
| Aspect ratio | 0.995 | 0.994 | 0.8% | 0.997 |
| Roundness  | 0.993 | 0.993 | 0.9%  | 0.999 |

*Lumen*

| Descriptor | Pearson r | CCC | Rel. error | Scale factor |
|------------|-----------|-----|------------|--------------|
| Area       | 0.955 | 0.892 | 14.0% | 1.110 |
| Perimeter  | 0.790 | 0.681 | 13.4% | 0.909 |
| Major axis | 0.941 | 0.842 | 9.1%  | 1.077 |
| Minor axis | 0.939 | 0.759 | 11.1% | 1.101 |
| Max Feret  | 0.924 | 0.917 | 5.1%  | 1.010 |
| Min Feret  | 0.935 | 0.915 | 5.3%  | 1.034 |
| Aspect ratio | 0.925 | 0.881 | 4.1% | 0.983 |
| Roundness  | 0.893 | 0.802 | 6.0%  | 0.954 |

The size bias (+6–14%) and the near-perfect agreement of the shape descriptors
(tubule AR/roundness CCC ≈ 0.99, scale factor ≈ 1.00) reproduce the in-domain pattern,
confirming the offset is a **domain-stable** scaling difference, not out-of-domain distortion.
Lumen perimeter is the only noisy metric (r 0.79), reflecting the scalloped, low-contrast
lumen boundary.

### Known biases

- Boundary accuracy vs. reference masks (validation, n = 49):
  - Tubule: ASSD 0.59 µm, HD95 1.54 µm, mean signed displacement +0.13 µm (Boundary-F1 @1.6 µm = 0.97).
  - Lumen: ASSD 1.47 µm, HD95 4.99 µm, mean signed displacement −0.05 µm (Boundary-F1 @1.6 µm = 0.75).

  The model reproduces the reference masks with sub-pixel mean boundary agreement; the
  external boundary displacement stays small (tubule +0.48 µm ≈ 1.5 px). The ~+12% / +8%
  (tubule / lumen) 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 it. It is systematic (r ≈ 0.99) and
  preserves shape (**tubule** AR/roundness CCC ≈ 0.99; lumen roundness lower, CCC 0.82).

- **Lumen perimeter**: good mean agreement in-domain (CCC 0.93, bias ≈ 0) but wide LoA
  (±110 µm), varying case by case with tracing convention.

### Failure mode

One badly-fixed external tubule was grossly under-segmented: the lumen was missed
entirely and only ~⅓ of the tubule area was captured (tubule boundary displaced
−10 µm inward, a clear outlier). The model's **status output flagged it as CHECK**
(1 of 9 flagged across the 129 images), so the failure is surfaced for review rather
than passed off as a valid measurement.

### Validated scope

Single-tubule crops at 0.32 µm/px, H&E staining. **Not validated** on multi-tubule fields or
other magnifications.

---

## Attribution

Dataset curated by Lucila Gotfryd (image acquisition, annotation, design of the segmentation
and morphometry approach, including the anatomically-motivated containment constraint and the
choice to enforce tubular connectivity). Model implementation carried out with AI coding
assistance ([ML Intern](https://github.com/huggingface/ml-intern)) under the author's direction.