[AAAI-26] Tuning Medical Foundation Models for Inner Ear Temporal CT Analysis with Plug-and-play Domain Knowledge Aggregator

Plug-and-Play Domain Knowledge Aggregator for inner ear CT analysis.

📜 Introduction

The CIED dataset is constructed by Visual and Intelligent Learning Lab, Tongji University and Department of Otolaryngology-Head and Neck Surgery, Shanghai Ninth People's Hospital, Shanghai Jiao Tong University School of Medicine.

Fig 1. The overall architecture of our pipeline. DKGT consists of two main components: Domain Knowledge Generator and Domain Knowledge Aggregator.

Training

Clone the code and follow the institution to install Merlin and CT-FM and download their pretrained models.

To train/evaluate DKGT on CT-FM on IEITD, do

cd CT-FM/
python ci_pmt_train.py --lr learning_rate --wd weight_decay --fold $fold

To train/evaluate DKGT on Merlin on IEITD, do

cd Merlin/documentation/
python ci_pmt_train.py --lr learning_rate --wd weight_decay --fold $fold

To visualize DKA on Merlin on IEITD, do

cd Merlin/documentation/
python visualize.py --fold $fold

TODO

Upload SAM-Med3D and model weights.

Dataset

The CIED is now available on Google Drive.

Reference

This code is based on Merlin and CT-FM.