Related Experiment Video
Updated: Jul 16, 2026
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
Metadata Conditioning via Feature-Wise Linear Modulation for Multi-Domain 3D Abdominal Segmentation: A Comparative
Juan F Garrido-Martínez1, Javier M Garrido-López1, Juan F Zapata-Pérez1
1Escuela Técnica Superior de Ingeniería Industrial, Campus Muralla del Mar, Universidad Politécnica de Cartagena, Member of European University of Technology EUT+, C/Doctor Fleming, s/n, 30202 Cartagena, Spain.
Feature-wise Linear Modulation (FiLM) metadata conditioning improved U-Net for 3D abdominal MRI segmentation, especially with heterogeneous data. However, FiLM did not consistently benefit nn-Net, highlighting architecture-dependent performance gains.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Automatic 3D abdominal organ segmentation in MRI is challenging due to data heterogeneity.
- Differences in acquisition protocols and imaging characteristics degrade model generalization.
- Metadata conditioning offers a potential solution to improve model adaptation.
Purpose of the Study:
- To investigate the efficacy of Feature-wise Linear Modulation (FiLM) for metadata conditioning in 3D segmentation networks.
- To compare the performance of U-Net and nnU-Net v2, with and without FiLM conditioning.
- To analyze the impact of FiLM on multi-domain robustness using CHAOS and AMOS datasets.
Main Methods:
- Implemented FiLM conditioning in U-Net and nnU-Net v2 architectures.
- Evaluated baseline and FiLM-conditioned models on CHAOS, AMOS, and combined datasets.
- Assessed segmentation performance using Dice, HD95, and ASD metrics.
- Analyzed the effect of post-processing on segmentation outputs.
Main Results:
- FiLM conditioning showed favorable trends in U-Net, particularly in combined datasets, improving Dice, HD95, and ASD.
- FiLM did not consistently benefit nnU-Net, and even reduced performance in some scenarios (e.g., AMOS).
- Post-processing had a greater impact on U-Net than nnU-Net, indicating nnU-Net's inherent stability.
Conclusions:
- FiLM can enhance generic architectures like U-Net for heterogeneous medical imaging segmentation tasks.
- The benefits of metadata conditioning are architecture-dependent and not universally applicable.
- nnU-Net v2 baseline demonstrated robust performance, requiring less benefit from additional conditioning mechanisms.