Depth-of-focus enhancement in optical coherence tomography via a cascaded image registration and fusion network for multi-focus imaging.

Abstract

Significance: Optical coherence tomography (OCT) is widely used in biomedical imaging, but its ability to clearly resolve fine structures is limited to a narrow depth of focus (DOF). This constraint restricts reliable visualization of tissue microstructures across extended depth ranges, making strategies that extend the DOF while preserving fine structural fidelity and image sharpness highly desirable.

Aim: We aim to enhance the effective DOF of OCT imaging while preserving fine structural details and image sharpness by developing a deep-learning-based reconstruction framework for multi-focus OCT data.

Approach: We developed a cascaded image registration and fusion network (CRFN) to process multi-focus OCT images acquired using a swept-source OCT system with dynamic focal modulation enabled by an electrically tunable lens. The proposed network consists of a registration module for spatial alignment of multi-focus images and a fusion module for focus map-guided reconstruction. CRFN operates in an unsupervised, training-free manner, in which the network parameters are optimized directly on the acquired multi-focus OCT images, without relying on large-scale pre-collected training datasets.

Results: Experiments conducted on ex vivo and in vivo specimens demonstrate that the proposed method improves image quality compared with unfused and conventional multi-focus fusion approaches. Quantitative evaluations on ex vivo specimens show maximum improvements of 8.29 dB in signal-to-noise ratio and 1.56 dB in contrast-to-noise ratio, compared with conventional methods. Moreover, the effective DOF is extended by a factor of $\sim 2$ , enabling clearer visualization of fine structural details across an enlarged depth range.

Conclusions: The proposed CRFN improves multi-focus OCT reconstruction quality and extends the effective DOF without increasing hardware complexity or relying on extensive training data, highlighting its robustness and potential generalizability for biomedical OCT imaging applications.