Restoration of out-of-focus region for photoacoustic endoscopic imaging enabled by convolutional neural networks

Photoacoustic endoscopy (PAE) enables high-resolution imaging of internal tissues but suffers from image degradation in out-of-focus regions, caused by laser beam divergence and reduced photoacoustic signal amplitude. This study investigates convolutional neural network (CNN) models for restoring out-of-focus PAE images. Four CNN models—feature fusion attention network (FFA), residual channel attention network (RCA), enhanced deep super resolution network (EDSR), and residual in residual dense block network (RRDB)—were compared with traditional methods, including Richardson–Lucy (RL) deconvolution and directional model-based deconvolution (DMB). CNN-based restoration of out-of-focus PAE images from phantom experiments using a 5-μm carbon fiber and 10, 20, and 40-μm tungsten wires, and a leaf phantom showed significant improvements in image quality. RRDB achieved the lowest relative size error (RSE) among all the tested algorithms, reducing RSE for out-of-focus PAE images of the 5-μm carbon fiber from 35.93% to 6.65%, the 10-μm tungsten wire from 29.97% to 4.89%, the 20-μm tungsten wire from 14.93% to 3.80%, the 40-μm tungsten wire from 6.15% to 3.86%, and the leaf phantom from 19.57% to 3.01%. The results demonstrate the potential of CNNs, especially RRDB, for effectively restoring out-of-focus PAE images.