ikchain/vet-dermatology-canine

Canine Dermatology Classifier — Howl Vision

EfficientNetV2-S for canine skin lesion classification (6 classes). Part of Howl Vision for the Gemma 4 Good Hackathon.

Metrics (held-out test, n=433)

Metric	Value
Accuracy	94.0% [91.5%, 95.8%] Wilson CI 95%
F1-macro	0.923

Details

• Architecture: timm/tf_efficientnetv2_s.in21k_ft_in1k, num_classes=6
• Input: 384x384 RGB, ImageNet normalization
• Classes: demodicosis, Dermatitis, Fungal_infections, Healthy, Hypersensitivity_Allergic_Dermatitis, ringworm
• Checkpoint: PyTorch .pt with model_state_dict

import timm, torch
from torchvision import transforms as T
from PIL import Image

model = timm.create_model("tf_efficientnetv2_s.in21k_ft_in1k", pretrained=False, num_classes=6)
ckpt = torch.load("vet_dermatology.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()