Development of a pipeline to diagnose prion disease by applying artificial intelligence on IHC-stained whole slide images

Prion disease is a neurodegenerative pathology characterized by deposits of the aberrant isoform of the prion protein known as scrapie prion protein (PrPsc) in the CNS. The PrPsc aggregates are highlighted by the immunohistochemical (IHC) staining with 12F10 antibody, and they constitute an essential feature to diagnose the disease. The development of specific scanners to digitalize slides recently made it possible to start a technological revolution in pathology since whole slide images (WSIs) can be analyzed with artificial intelligence. In this work, it was developed a pipeline to perform the diagnosis of prion disease from WSIs referring to tissue samples extracted from the cerebellar or the occipital cortex. The aim was to offer a second quantitative opinion to the pathologist to minimize errors and permit him to work efficiently, quickly, and qualitatively. The key element of the pipeline is an algorithm to recognize histopathological patterns. Two different strategies were tested: a deep learning approach based on the use of a DNN for segmentation with an architecture of the type of vision transformer, and a machine learning approach, which involved training a traditional classifier such as ANN or SVM to address the task. Both strategies required the extraction of tiles from the WSIs, which was performed using manual annotations to select informative tissue regions and exclude artifacts. Patching was then performed on the tiles to obtain the input images for the DNN. Traditional classifiers required an additional step: patches were converted into single-channel images, and texture analysis was used to extract features to construct the dataset for training. Once the models were trained, they were both evaluated on the entire tiles to compare the two approaches. A diagnostic criterium based on the percentage of positive pixels was applied to perform the final diagnosis. The two models gave good results concerning both Construction and Test sets. Segmentation performs pattern recognition more accurately. Another difference between the two approaches lies in the application time since feature extraction requires a large amount of time. However, in terms of diagnostic performance, the two approaches were comparable. A potential earlier step of the pipeline was tested, which consists of performing stain transfer to produce the IHC stain from the WSI stained with H&E. A Pix2Pix Generative Adversarial Network (GAN) was trained on paired images. To obtain the dataset for training, WSIs with different stains were registered using the ORB detector, then some tiles were extracted from both, and a second registration was performed to obtain a pixel-level overlap. Virtual staining would make it possible to skip the expensive and time-consuming immunostaining process, and it would solve the problem of staining variability due to the operator and the instrumentation, which strongly affected the provided dataset and made this work more challenging.