Registration of differently stained histopathology images for tissue evaluation

Histopathology image registration is crucial for appropriate tissue analysis, especially for the diagnosis of cancer. The presence of staining variation among these whole-slide images leads to several challenges because the stain differs greatly across various tissue types. There are multiple registration techniques, including intensity-based, feature-based, and deep learning-based methods. Although these methods employ different working mechanisms, they all face several challenges, including poor generalization due to staining discrepancies. To address this problem, this study proposed a highly optimized and robust landmark-based registration technique. The technique utilizes anatomical landmarks and employs thin-plate spline interpolation to achieve precise non-rigid alignment, effectively addressing both tissue deformation and staining variability. Moreover, stain normalization was applied using the deep learning-based model StainNet to evaluate how changes in color distribution in images affect the registration approach. Furthermore, the developed approach was comprehensively compared with two well-known intensity-based methods, Elastix and ANTs, under both normalized and raw image conditions. The results from the experiments showed that landmark-based registration was more effective under both conditions in dealing with staining changes than Elastix and ANTs. While Elastix significantly struggled with normalized images due to its reliance on pixel intensities, ANTs showed increased strength after stain normalization; however, they did not perform as well as the method based on landmarks. Although, landmark-based registration is reliable as it is not influenced by varying staining but it is still challenging to annotate the landmarks with significant accuracy that can be improved in future techniques by using automated landmark annotations.