Missing SLO scans no longer mean zero-filling: conditional diffusion recovers a proxy representation and lifts postoperative vision prediction accuracy to 0.9714

A missing SLO modality is no longer simply discarded or zero-filled; the model estimates a sample-specific proxy representation in latent feature space and uses it as a substitute input for postoperative best-corrected visual acuity prediction. Earlier missing-modality methods mainly reduced the negative influence of missing inputs during fusion or learning, or learned modality-invariant representations, without explicitly recovering a task-relevant representation for the missing SLO; the compared UML backbone baseline handles missing SLO through mask-based fusion. Under the same data split, prediction backbone and evaluation metrics, the diffusion recovery method raised accuracy from the UML baseline's 0.9625 to 0.9714 and lowered mean absolute error from 0.0457 to 0.0423, achieving the highest accuracy and lowest mean absolute error among comparisons with MAF-Net, CTT-Net, MMIN and IML-Net.

Recovery does not generate raw SLO images; OCT image features and clinical text features are first fused into a conditional representation, which then guides progressive denoising from noise into an SLO proxy representation. The conditional diffusion model here performs reverse denoising in feature space, trained with a noise prediction loss, a representation reconstruction loss and a cosine direction constraint so that the recovered representation stays close to the real SLO representation in value and overall direction. The recovery module is pretrained only on modality-complete samples with a 100-step DDPM, and the target SLO representation has 49 tokens with 768 feature dimensions; on 101 validation samples with available SLO, recovered representations reached a mean cosine similarity of 0.9583 and an MSE of 0.00118 against real SLO features.

How conditional information is organized directly shapes downstream prediction, and fusing OCT with text to build the condition works best. OCT-only conditioning, text-only conditioning and simple concatenation of OCT and text all fell short of the fused condition; text-only conditioning dropped accuracy to 0.9442 and raised mean absolute error to 0.0507. In the ablation, fused conditioning reached 0.9714 accuracy and 0.0423 mean absolute error, but its test loss of 2.0068 was not the lowest across all metrics.

Results come from a single-center set of 2,141 preoperative multimodal cataract samples with a 58.80% SLO missing rate; missing rates differ naturally across splits, with 66.05% in the test set. The metrics are accuracy under an absolute error within 0.1 LogMAR and mean absolute error, where the 0.1 LogMAR threshold corresponds approximately to one line on a standard LogMAR chart and is described as an interpretable criterion rather than a validated threshold for clinical decision-making.

The authors note limits from a single-center dataset, naturally varying SLO missing rates across splits and the absence of external validation, and call for further evaluation on independent clinical cohorts and under controlled missingness settings. Recovered representations also show a more concentrated distribution than real SLO features, so high representation similarity does not necessarily preserve full inter-sample variation.