Research map / Label-efficient and robust learning / Recognition and long-tail
Experimental Design for Multi-Channel Imaging via Task-Driven Feature Selection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TADRED, a task-driven feature-selection method for experimental design in multi-channel imaging that jointly trains a scoring network and task network while iteratively subsampling imaging channels via a four-phase RFE-style procedure.
Problem
Choosing a compact set of image channels (acquisition contrasts or spectral filters) that best supports a user-specified imaging task, where classical Fisher-matrix and Bayesian experimental design are model-specific and intractable in high dimensions.
Approach
Recasts experimental design as supervised feature selection over densely sampled channels: a scoring network and task network are trained together, with stepwise recursive feature elimination and a four-phase schedule that gradually moves from sample-dependent to sample-independent scores and from soft to binary masks.
Key results
- On VERDICT-MRI with C=110 of 220 designs, TADRED achieves MSE 1.03e-2 versus 1.06e-2 for SSEFS and 2.22e-2 for FIRDL, with the gap widening at smaller C (2.64e-2 vs 4.58e-2 at C=14).
- On AVIRIS remote sensing reconstruction at C=14, TADRED achieves MSE 5.80 versus 8.27 for random and 10.3 for FIRDL, and it outperforms SSEFS at every subsample size.
- On the MUDI challenge, TADRED reduces MSE to 0.22 at C=500 versus 0.49 for PROSUB, and it matches or beats PROSUB at all tested sizes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TADRED channel selection instead of SSEFS | VERDICT-MRI parameter estimation · VERDICT-MRI, C=14 | MSE × 10²: 4.58 -> 2.64 | Helped |
| Use TADRED reconstruction instead of PROSUB | MUDI MRI reconstruction · MUDI, C=500 | MSE: 0.49 -> 0.22 | Helped |
| Use TADRED channel selection instead of random selection | AVIRIS Indian Pine reconstruction · AVIRIS Indian Pine, C=110 | MSE: 1.81 -> 0.87 | Helped |
| Use TADRED channel selection instead of the published baseline | Tissue oxygenation estimation · Tissue oxygenation, HbO₂/Hb abundance, C=6 | RMSE × 10²: 4.54 -> 2.80 | Helped |
| Use TADRED design instead of classical Fisher-matrix design | NODDI parameter estimation · NODDI, C=99 | MSE × 10²: 8.00 -> 4.51 | Helped |
| Remove the scoring network (ablation) | TADRED for VERDICT-MRI · VERDICT-MRI, C=110 | MSE × 10²: 1.03 -> 7.23 | Hurt |
| Remove the scoring network (ablation) | TADRED for VERDICT-MRI · VERDICT-MRI, C=55 | MSE × 10²: 1.18 -> 10.7 | Hurt |
| Remove iterative subsampling (ablation) | TADRED for VERDICT-MRI · VERDICT-MRI, C=14 | MSE × 10²: 2.64 -> 2.80 | Hurt |
| Use data mean rather than data median for feature fill (ablation) | TADRED for VERDICT-MRI · VERDICT-MRI, C=14 | MSE × 10²: 2.55 -> 2.80 | Hurt |
| Use zeros rather than data median for feature fill (ablation) | TADRED for VERDICT-MRI · VERDICT-MRI, C=14 | MSE × 10²: 2.55 -> 2.51 | Helped |
| Replace learned feature scores with random scores (ablation) | TADRED channel selection · VERDICT-MRI experimental settings | MSE | Hurt |
Surprising findings
- Random channel selection often outperforms SSEFS when no informative correlated feature subset exists, which the authors attribute to the experimental design setting.
- A variant of TADRED that uses random scores in place of learned scores performs more than 10% worse, indicating the scoring network contributes to performance.
Limitations
- Iterative subsampling may decrease the performance upper bound because optimal feature sets for successive sizes need not be nested.
- TADRED is likely to underperform SSEFS and FIRDL on standard feature selection problems with small discriminative subsets among many uninformative features.
- Only pixel-wise independent tasks with fully-connected networks are considered; equal channel costs are assumed.
Experimental setup
Datasets: VERDICT-MRI simulations NODDI simulations MUDI Multi-Diffusion challenge data HCP diffusion MRI data AVIRIS Indian Pine hyperspectral data hyperspectral tissue oxygenation simulations
Benchmarks: VERDICT-MRI parameter estimation MUDI Multi-Diffusion challenge HCP downstream DTI/DKI/MSDKI metrics AVIRIS Indian Pine reconstruction hyperspectral tissue oxygenation estimation NODDI parameter estimation
Metrics: MSE RMSE Jaccard index standard deviation across seeds
Baselines: Fisher-matrix experimental design PROSUB SSEFS FIRDL Random supervised feature selection
Builds on: FIRDL SSEFS PROSUB recursive feature elimination Fisher information matrix experimental design
Closest papers by meaning
- AeSPa : Attention-guided Self-supervised Parallel Imaging for MRI Reconstruction (CVPR 2025 · similarity 0.83)
- Information-Driven Design of Imaging Systems (NeurIPS 2025 · similarity 0.82)
- Enhancing Feature Diversity Boosts Channel-Adaptive Vision Transformers (NeurIPS 2024 · similarity 0.82)
- FactorizePhys: Matrix Factorization for Multidimensional Attention in Remote Physiological Sensing (NeurIPS 2024 · similarity 0.82)
- ChAda-ViT : Channel Adaptive Attention for Joint Representation Learning of Heterogeneous Microscopy Images (CVPR 2024 · similarity 0.81)
- Hybrid-Supervised Dual-Search: Leveraging Automatic Learning for Loss-Free Multi-Exposure Image Fusion (AAAI 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.