GJFocuser: a Gaussian difference and joint learning-based autofocus method for whole slide imaging

Whole slide imaging (WSI) provides tissue visualization at the cellular level, thereby enhancing the effectiveness of computer-aided diagnostic systems. High-precision autofocusing methods are essential for ensuring the quality of WSI. However, the accuracy of existing autofocusing techniques can be notably affected by variations in staining and sample heterogeneity, particularly without the addition of extra hardware. This study proposes a robust autofocusing method based on the difference between Gaussians (DoG) and joint learning. The DoG emphasizes image edge information that is closely related to focal distance, thereby mitigating the influence of staining variations. The joint learning framework constrains the network's sensitivity to defocus distance, effectively addressing the impact of the differences in sample morphology. We first conduct comparative experiments on public datasets against state-of-the-art methods, with results indicating that our approach achieves cutting-edge performance. Subsequently, we apply this method in a low-cost digital microscopy system, showcasing its effectiveness and versatility in practical scenarios.