Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
OCTDiff: Bridged Diffusion Model for Portable OCT Super-Resolution and Enhancement
Authors: Ye Tian, Angela McCarthy, Gabriel Gomide, Nancy Liddle, Jedrzej Golebka, Royce Chen, Jeff Liebmann, Kaveri Thakoor
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | OCTDiff outperforms convolutional baselines and standard DDPMs, achieving state-of-the-art performance on clinical portable OCT datasets. Our model and its downstream applications have the potential to generalize to other medical imaging modalities and revolutionize the current workflow of ophthalmic diagnostics. We present quantitative and qualitative results of our OCTDiff against baseline models. We conduct ablation studies to analyze the effects of the ANA and MSCA modules and the quality-score informed loss on model performance, complexity, and training efficiency, including different cross-attention types and the exponential decay rate α in ANA. |
| Researcher Affiliation | Academia | 1Department of Biomedical Engineering, Columbia University, New York, NY, USA 2Department of Computer Science, Columbia University, New York, NY, USA 3Department of Ophthalmology, Columbia University Irving Medical Center, New York, NY, USA EMAIL |
| Pseudocode | Yes | Algorithm 1 Adaptive Noise Aggregation (ANA) Require: High-resolution input ˆx T = y, total steps T Ensure: Super-resolved output ˆx0 1: for t = T 1 to 0 do 2: ˆϵt εθ(ˆxt+1, t) 3: if t > T/2 then 4: ˆϵt Refine(ˆϵt, Ldenoise) 5: end if 6: ϵt 1 Zt PT 1 τ=t exp( α(τ t)) ˆϵτ 7: ˆxt 0 Reconstruct(ˆxt+1, ϵt, t) 8: ˆxt µθ(ˆxt+1, ˆxt 0, t) 9: end for 10: return ˆx0 |
| Open Source Code | Yes | The code is available at https://github.com/AI4VSLab/OCTDiff. |
| Open Datasets | No | Due to the lack of publicly available low-resolution OCT datasets, we collected our own real-world dataset with the Philophos KUOS-O100 portable OCT device. The complete clinical dataset is currently under IRB restrictions and ongoing collection, and will be made publicly available in a separate publication. |
| Dataset Splits | Yes | We performed 5-fold cross-validation for each model-dataset pair, with the results shown in Table 6, and we conducted Mann Whitney U tests [62] to assess statistical significance. |
| Hardware Specification | Yes | Training is conducted on a Lambda Labs Vector server equipped with two NVIDIA A6000 GPUs, requiring approximately 48 hours to complete from scratch with input images resized to 256 × 256 pixels and a total diffusion time step T = 1000. |
| Software Dependencies | No | The paper mentions 'Im Fusion software [49]' for affine registration, but does not provide specific version numbers for this or any other key software components used in the experiments. |
| Experiment Setup | Yes | Training is conducted on a Lambda Labs Vector server equipped with two NVIDIA A6000 GPUs, requiring approximately 48 hours to complete from scratch with input images resized to 256 × 256 pixels and a total diffusion time step T = 1000. The scalar α > 0 controls how fast the weight decays over time. We discuss the impact of different α values in an ablation study (Section 4.2). This trade-off suggests that a moderate α achieves the best balance, and thus we selected α = 0.3 for our experiments. Here, γ < 0 is a focusing parameter that controls the degree to which high-quality samples are prioritized, usually being set in [-2, -1]. |