CogGen: Cognitive-Load-Informed Fully Unsupervised Deep Generative Modeling for Compressively Sampled MRI Reconstruction

TL;DR

CogGen is proposed, a cognitive-load-informed FU-DGM that casts CS-MRI as staged inversion and regulates task-side"cognitive load" by progressively scheduling intrinsic difficulty and extraneous interference and improves fidelity and convergence over strong unsupervised baselines and competitive supervised pipelines.

Abstract

Fully unsupervised deep generative modeling (FU-DGM) is promising for compressively sampled MRI (CS-MRI) when training data or compute are limited. Classical FU-DGMs such as DIP and INR rely on architectural priors, but the ill-conditioned inverse problem often demands many iterations and easily overfits measurement noise. We propose CogGen, a cognitive-load-informed FU-DGM that casts CS-MRI as staged inversion and regulates task-side "cognitive load" by progressively scheduling intrinsic difficulty and extraneous interference. CogGen replaces uniform data fitting with an easy-to-hard k-space weighting/selection strategy: early iterations emphasize low-frequency, high-SNR, structure-dominant samples, while higher-frequency or noise-dominated measurements are introduced later. We realize this schedule via self-paced curriculum learning with complementary student-mode (what the model can currently learn) and teacher-mode (what it should follow) criteria, supporting both soft weighting and hard selection. Experiments and analysis show that CogGen-DIP and CogGen-INR improve fidelity and convergence over strong unsupervised baselines and competitive supervised pipelines.

Figures (6)

• Figure 1: Progressive k-space sampling strategies in CogGen framework, showing the evolution of sample selection across five training curricula ($C_1$-$C_5$). Top: Data$\#1$ (2D, AF = 8); Bottom: Data$\#2$ (1D, AF = 6). The $C_5$ refers to the actual undersampling k-space measurement.
• Figure 2: Data$\#1$ reconstruction results at AF = 8. (a)-(i): DIP-TV, BM3D-FISTA, DISCUS, SSDU, aSeq-DIP, Hash-INR-Elastic, MoDL, CogGen-DIP and CogGen-INR. The corresponding error maps are presented in the bottom row, alongside the GT and downsampling mask displayed at the far left. Our approaches achieve superior accuracy, surpassing all competing methods.
• Figure 3: Data$\#2$ reconstruction results at AF = 6. (a)-(i): DIP-TV, BM3D-FISTA, DISCUS, SSDU, aSeq-DIP, Hash-INR-Elastic, MoDL, CogGen-DIP and CogGen-INR. The corresponding error maps are presented in the bottom row, alongside the GT and downsampling mask displayed at the far left. Our approaches achieve superior accuracy, surpassing all competing methods.
• Figure 4: Benefits of the CogGen framework on Data$\#1$ (2D, AF = 8) and Data$\#2$ (1D, AF = 6). Across both DIP and INR paradigms, incorporating CogGen yields marked gains in reconstruction accuracy and convergence efficiency. (a-d): DIP, INR, CogGen-DIP and CogGen-INR.
• Figure 5: The effect of curriculum size on Data$\#3$ at AF = 10. Performance improves up to $C_4$, beyond which further stage increases lead to degradation, indicating that the curriculum size must be carefully tuned to optimize the learning progression.

Theorems & Definitions (5)

• Definition 1: Deep Network Prior jagatap2019phasesaragadam2024deeptensor
• Lemma 6.1: Stage-wise linear convergence
• proof
• Theorem 6.1: Accelerated convergence of CogGen
• proof