Tubules_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)

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.

---

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

Tubule

Lumen

---

3. Data and training

---

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):

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):

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

Tubule

Lumen

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) under the author's direction.