Abstract:
Human pathologists perform pathological diagnosis by analyzing the stained specimen on a glass slide under a microscope. But in recent years, multiple scanners have been introduced which can capture the entire stained glass slide and save it as a digital image called Whole Slide Image(WSI). This results in the digitization of patient specimens which later can easily be globally distributed for research, teaching and diagnosis purposes. This adoption of digital pathology has the potential to further automate the process of pathologists based on deep learning algorithms resulting in reduced turnaround time, decreasing the workload of pathologists and further standardizing clinical practices. Digital pathology using deep learning approaches are often general image recognition techniques showing significant improvements on visual understanding. However, whole slide images have billions of pixels containing various types of artifacts and high morphological heterogeneity, which makes this image analysis task particularly challenging. These non-trivial conditions impede the conventional deep learning techniques and thereby need some special processing techniques and some form of dimensionality reduction to the images. This study aims to develop a convolutional neural network based algorithm for identification of viable areas in the whole slide image by dividing it into smaller patches. We can call this a divide and conquer approach where we basically divide WSI into smaller images with viable/not viable classification and then at the end we combine all these patches to get the viable area in the whole slide image. Results demonstrate superior performance for classification of patches identifying optimal area in whole slide image which further can be used for cell morphological analysis which is important for the diagnosis of illness of an individual for various types of diseases.
