Metaseg Finer 2d 5cls

--- license: apache-2.0 language: en library_name: pytorch tags: - medical-imaging - brain-mri - segmentation - implicit-neural-representation - meta-learning - finer - parameter-efficient datasets: - oasis metrics: - dice pipeline_tag: image-segmentation --- # MetaSeg-FINER 2D 5-class (OASIS, retuned `omega_0=15`) Meta-learned **FINER implicit neural network** (Liu et al., CVPR 2024) for 5-class brain MRI segmentation on neurite-OASIS coronal slices. Plugged into the MetaSeg recipe with a single hyperparameter retune (`omega_0=15`, half the FINER paper default). **Slightly exceeds SIREN** at identical parameter count. ## Performance | Metric | Value | Notes | |---|---:|---| | Mean Dice (OASIS test split, n=80) | **0.929 ± 0.012** | +0.004 over SIREN baseline (0.925) at identical params | | INR parameters | 83 K | Same as SIREN reference | | `omega_0` | **15.0** | Retuned: FINER paper default is 30 | | `first_bias_scale` | 10.0 | FINER paper default | | `scale_req_grad` | False | Detach scale term for MAML grad stability | ## The retune FINER's paper-default hyperparameters lose to SIREN by ~0.04 Dice inside the MetaSeg recipe (FINER default → 0.886, SIREN → 0.925). The activation `sin(omega_0 * (|x| + 1) * x)` already amplifies frequency via the `|x|+1` factor; layering SIREN's default `omega_0=30` on top destabilises second-order MAML gradients. **Halving `omega_0` to 15 closes the gap and slightly exceeds SIREN**, and the same retune transfers cleanly to the 25-class task (`metaseg-finer-2d-25cls` would close the gap there too, given the same fix). This is evidence that **the MetaSeg recipe is genuinely INR-agnostic** — but only if you treat backbone hyperparameters as part of the recipe, not as fixed defaults inherited from the upstream signal-fitting use case. ## Usage ```python from inr_brain_seg import InrBrainSeg model = InrBrainSeg.from_pretrained("basimazam/metaseg-finer-2d-5cls") mask = model.segment("path/to/T1.nii.gz") ``` ## Training details Same as the SIREN 2D-5cls model except: * **Backbone**: FINER with `omega_0=15`, `first_bias_scale=10`, `scale_req_grad=False`. ## Citation ```bibtex @inproceedings{liu2024finer, title = {{FINER}: Flexible Spectral-bias Tuning in Implicit Neural Representation by Variable-periodic Activation Functions}, author = {Liu, Zhen and others}, booktitle = {CVPR}, year = {2024} } ``` #...

Source context: 5 downloads · 0 likes · Pipeline image-segmentation · Library pytorch · Repo basimazam/metaseg-finer-2d-5cls