AFA-LoRA

Amplified fold-averaged low-rank adaptation with family-balanced logit fusion, for frame-level detection of neoplasia in Barrett's oesophagus. Final weights of Team GenMI's submission to the RARE 2026 challenge (MICCAI 2026).

Code: github.com/adinathdukre/AFA-LoRA

Model

25 binary frame classifiers, each a timm encoder with a linear head, initialised from the public GastroNet-5M self-supervised checkpoints.

Members	Encoder	Input	Adaptation	Family
15	DINOv2 ViT-B/14, 4 registers	336x336	LoRA r16, merged (x1.5, fold-averaged)	dinov2_vitb
5	ResNet-50 (DINO)	384x384	full fine-tune	resnet50_dino
5	ResNet-50 (MoCo v2)	384x384	full fine-tune	resnet50_mocov2

The LoRA adapters are already merged, so every member is a plain timm state dict. Each member's logit is standardised with the constants in ensemble.json, averaged within family, averaged across families, and mapped to (0, 1) by 0.5 + arctan(x) / pi.

Files

ensemble.json                     fusion rule, member order, per-member temperature, bias and family
<run>_fold_<k>/meta.json          timm architecture, input size, family
<run>_fold_<k>/model.safetensors  weights

Usage

git clone https://github.com/adinathdukre/AFA-LoRA
cd AFA-LoRA
pip install -r requirements.txt huggingface_hub
huggingface-cli download adidukre/AFA-LoRA --local-dir /path/to/AFA-LoRA-weights

import sys

import numpy as np
import torch

sys.path.insert(0, "submission")
from ensemble import RareEnsemble

ensemble = RareEnsemble.from_directory("/path/to/AFA-LoRA-weights", torch.device("cuda"))
scores = ensemble.predict(np.zeros((4, 512, 640, 3), dtype=np.uint8), batch_size=32)

Input is a stack of uint8 RGB frames of shape (N, H, W, 3). The output is a ranking score, not a calibrated probability.

Training data

Only the official RARE release: 3,095 frames (158 neoplasia) from two centres.

Limitations

Research use only. Not a medical device and not validated for clinical decisions.

Licence

MIT. The GastroNet-5M checkpoints and the RARE data are subject to their own terms.