ISIC Skin Lesion Segmentation — SegFormer (MiT-B0)

Pixel-level skin lesion segmentation model fine-tuned on ISIC 2018 Task 1.

Model Details

Architecture: SegFormer (NVIDIA MiT-B0)
Task: Binary medical image segmentation (lesion vs background)
Framework: PyTorch + Hugging Face Transformers
Input size: 256 × 256
Output: Lesion probability mask
Parameters: ~3.7M

Performance (ISIC 2018)

Metric	Score
Test Dice	90.32%
Test IoU	82.50%

Compared against a from-scratch U-Net baseline (85.26% Dice).

Files

best_segformer_isic2018.pth — fine-tuned weights

How to Load

from transformers import SegformerForSemanticSegmentation
from huggingface_hub import hf_hub_download
import torch

model = SegformerForSemanticSegmentation.from_pretrained(
    "nvidia/mit-b0",
    num_labels=1,
    ignore_mismatched_sizes=True
)

path = hf_hub_download(
    "mahnoor-2722/isic-segformer",
    "best_segformer_isic2018.pth"
)
model.load_state_dict(torch.load(path, map_location="cpu"))
model.eval()

Downloads last month: -; Downloads are not tracked for this model. How to track