Multimodal imaging has become an essential tool in clinical ophthalmology, offering complementary perspectives for disease diagnosis. However, current automated diagnostic approaches often fail to fully exploit the rich, complementary information provided by different imaging modalities. In this paper, to advance automated ophthalmic disease diagnosis through effective multimodal data integration, we propose the OphFusionNet, a novel multimodal learning framework based on uncertainty-driven multi-scale multimodal feature fusion. Drawing inspiration from clinical observations that ophthalmic lesions often appear at multiple spatial scales, we design a multi-scale feature fusion module with sparse self-attention (MSFF-SSA). This module captures hierarchical representations while suppressing redundancy, thereby enhancing both the expressiveness and efficiency of the extracted features. To further improve multimodal fusion, we introduce an uncertainty-aware multimodal fusion module with a game-theoretic selection strategy (UMF-GTSS). This component estimates the uncertainty associated with different features and adaptively weights them based on their relative reliability, yielding more robust and trustworthy diagnostic outcomes. To mitigate the tendency to over-rely on dominant modalities and underutilize the informative potential of subordinate ones, we propose a modality distillation strategy (MDS), which leverages multimodal features to guide and refine the learning of single-modal representations. This strategy enhances generalization and boosts the discriminative capacity of each individual modality. The OphFusionNet was evaluated on four publicly available ophthalmic datasets. Extensive experiments demonstrate that our approach achieves superior multimodal integration, resulting in state-of-the-art (SOTA) performance in multimodal ophthalmic disease diagnosis. The code will be available at: https://github.com/wb66715/OphFusionNet.

Contrastive learning, which aims to capture general representation from unlabeled images to initialize the medical analysis models, has been proven effective in alleviating the high demand for expensive annotations. Current methods mainly focus on instance-wise comparisons to learn the global discriminative features, however, pretermitting the local details to distinguish tiny anatomical structures, lesions, and tissues. To address this challenge, in this paper, we propose a general unsupervised representation learning framework, named local discrimination (LD), to learn local discriminative features for medical images by closely embedding semantically similar pixels and identifying regions of similar structures across different images. Specifically, this model is equipped with an embedding module for pixel-wise embedding and a clustering module for generating segmentation. And these two modules are unified by optimizing our novel region discrimination loss function in a mutually beneficial mechanism, which enables our model to reflect structure information as well as measure pixel-wise and region-wise similarity. Furthermore, based on LD, we propose a center-sensitive one-shot landmark localization algorithm and a shape-guided cross-modality segmentation model to foster the generalizability of our model. When transferred to downstream tasks, the learned representation by our method shows a better generalization, outperforming representation from 18 state-of-the-art (SOTA) methods and winning 9 out of all 12 downstream tasks. Especially for the challenging lesion segmentation tasks, the proposed method achieves significantly better performance.

Synthetic Aperture Radar (SAR) imaging can suffer distortion in the presence of phase errors in the acquired signal data caused by the non-ideal platform motion trajectory. Autofocus algorithms are used to remove the undesired phase error through signal processing techniques. Multi-channel convolution model of the SAR autofocus problem is established on the condition that phase error is range-independent. Then a new non-iterative autofocus algorithm is proposed based on cross correlation approach. The proposed autofocus algorithm can directly acquire the ideal SAR echo data without the estimation of phase error. Experimental results using the real SAR images demonstrate the validity of the proposed autofocus algorithm.