Despite rapid advancements in medical research, the benchmark for cancer diagnosis remains the histopathologic diagnosis. Histopathological images are microscopic images of the tissues used in disease analysis. The nature of histopathological images and the drastic rise in work, therefore, make this job lengthy and the findings may be subject to the pathologist’s subjectivity. Therefore, the production of automated and precise methods of histopathological image analysis is crucial for the area (Sudharshan et al., 2019). Cell-level histopathological image analysis Histopathological image analysis is widely used for cancer grading. Compared to mammography, CT and others, histopathology slides provide more comprehensive information for diagnosis and the diseases are analyzed by detecting tissue and cells in lesions (Gurcan et al., 2009). On the other hand an invasive biopsy is necessary, which is often tried to be avoided. CBIR systems are commonly employed to analyze histopathological images (Doyle et al., 2007; Caicedo et al., 2011; 2009). In CBIR systems, the returned visually similar images can be used to identify and classify the query images (e.g., classifying them as benign or malignant), and further assist pathologists to describe the tissue samples.
Hashing methods were first employed by Zhang et al. (2014); 2015c) to tackle large histopathological image databases for CBIR. They designed a comprehensive CBIR framework to analyze histopathological images by leveraging high-dimensional texture features and kernelized hashing with supervised information. In the experiment, this hashing method demonstrated significant improvement in histopathological image classification and retrieval tasks. Compared to methods such as SVM (Caicedo et al., 2009), Adaboost (Doyle et al., 2012), KNN (Tabesh et al., 2007), and Graph Embedding (Basavanhally et al., 2010), its accuracy was 5 to 10 percent higher. The method achieved histopathological retrieval for 700 − 900 images within 0.01 seconds (3121 images in the database), which is 1000 times faster than the given baseline. Fig. 8 illustrates four queries (two benign images, two malignant images) and their corresponding top five retrieval results based on this hashing-based CBIR framework. Despite the difference between benign and malignant images being subtle, the proposed method is effective to retrieve images in the same category. The authors extended the CBIR system for more accurate diagnosis by examining the cells in histopathology images (Zhang et al., 2015d). As each histopathology image usually includes thousands of cells, examining every cell by traditional retrieval methods is almost impossible when the image databases are large. Thus, a hashing-based framework is proposed that enables cell-level analysis in real-time with high accuracy, i.e., indexing 96, 000 cells within 1.68 seconds (the whole database includes 484,136 cells), and achieving 87.3% accuracy for the classification of histopathology lung images (i.e., two types of lung cancers, adenocarcinoma or squamous carcinoma).