Paper Review — verify claims with raw data

Concise critical evaluation (evidence-based)

• Strength — sensible architecture choice: Object-centric YOLOv8 segmentation matches the problem of sparse, discrete haustoria and organelles better than dense pixel-wise methods for these use-cases; authors justify choice and report strong nucleus metrics (>0.95) and practical runtime/annotation compromises Architecture rationale & nucleus metrics
• Strength — transfer learning & interpretability: Transfer learning (HFinder-2) improved recall and produced more spatially focused model attributions (occlusion analysis), suggesting improved confidence calibration rather than only incremental metric gains Transfer learning results
• Weakness — reliance on 2D single optical sections: Many analyses used single optical sections (not volumetric), which inflates ambiguity for hyphae crossing focal planes (authors note hyphae precision <0.50 and discuss need for 3D extension) — a real limitation for morphology/dynamics interpretation 2D limitation & hyphae precision
• Weakness — semi-supervised threshold-derived masks: Several large annotation pools (e.g., 10k+ chloroplast masks) were created semi-automatically via thresholding and not exhaustively hand-curated, causing merged-instance errors and lowering chloroplast precision (~0.7); this limits ground-truth fidelity and complicates error attribution (annotation noise vs model error) Semi-supervised annotation caveat
• Blindspot — domain shift across microscopes/species: Although transfer learning was effective within tested labs and microscopes, broader domain-shift robustness (different optics, staining, species, soil backgrounds) remains unproven and is correctly highlighted by authors as a limitation Domain-shift limitation
• Reproducibility & FAIRness: Good — authors released pretrained models and data on Zenodo and described preprocessing parameters (epsilon_rel, min_points, min_area, target_points), increasing reproducibility; however, per-instance IoU lists and full annotation manifests should be published to maximize reusability (authors reference Supporting Table 1 on Zenodo) Data & reproducibility statements

What would falsify the main claims?

1. Independent application of HFinder (or provided models) to fresh confocal datasets from different labs (distinct microscopes, fluorophores, and species) yields substantially lower recall/IoU than reported (<0.6 recall, median IoU <0.5) after minimal fine-tuning — would challenge claims of robustness and transferability Test of generalization
2. Full 3D volumetric imaging and annotation showing that many single-section 'haustoria' detections are mis-assigned across Z leading to inflated IoU in 2D — would indicate 2D approach misrepresents 3D morphology (authors recommend 3D expansion as future work) Need for 3D validation

Practical suggestions to strengthen/extend the work

• Release a per-instance annotated mask table and IoU/score CSV for each test image (enables exact replication of reported histograms and secondary analyses).
• Publish a small 3D (Z-stack) annotated subset and benchmark HFinder (2D vs 3D U-Net or 3D detection architectures) to quantify 2D-to-3D failure modes.
• Provide domain-adaptation recipes (contrast/PSF augmentation, stain transforms) and an evaluation on at least two independent external lab datasets to better measure generalization.
• Where semi-supervised masks were used, provide a curated subset of fully hand-annotated organelle instances to estimate annotation noise bias and correct precision/recall accordingly.

Concrete, testable follow-up experiments

1. Cross-lab blind benchmark: provide the pretrained HFinder-4 model to three independent labs (different confocal systems, fluorophores) and request predictions on a held-out set; compare IoU/precision/recall to reported metrics to quantify domain-shift impact.
2. 3D growth dynamics study: collect time-lapse Z-stacks of early haustorium formation and test whether a 2D HFinder pipeline systematically under- or over-estimates haustorial initiation events versus a 3D segmentation baseline.