Detection of liver fibrosis on slide-level labeled unstained biopsies by quantitative phase imaging and multiple instance learning

Abstract

In recent years, advances in artificial intelligence (AI) have had a great impact in medicine, particularly in diagnostic fields such as radiology and pathology. In this context, the recognition and detection of important patterns in medical data can be made with AI, which can improve diagnostic processes, automate and make them more accurate and efficient, as well as increase sensitivity and consistency. One area where AI has great potential is digital pathology (DP), which is increasingly replacing traditional histopathological examination. In DP, tissue slides are turned into high-quality digital images and it is very convenient to perform detection, segmentation and classification tasks using AI on these data. On the other hand, biopsy specimens are stained for imaging with contrast in conventional pathology and current digital pathology. However, these methods have disadvantages such as staining errors, color differences due to different stain reagent manufacturers and different staining conditions. Quantitative phase imaging (QPI) is a very suitable method to overcome these disadvantages. Biological samples used in classical microscope examination are very thin and poorly absorb and scatter light. Therefore, no contrast occurs during imaging without staining, ie they are transparent. Whereas conventional bright field imaging measures only the amplitude, OPI allows measurement of both amplitude and phase information of the sample. Thus, transparent samples can be imaged with high contrast via QPI without any staining. QPI has great potential for use in cell biology, imaging of red blood cells, and diagnosis of diseases by histopathological examination, such as cancer. Fourier Ptychography (FP) microscopy is a computational imaging technique that can perform QPI. FP combines two classical optical techniques, Phase Retrieval and Aperture Synthesis, to obtain a complex image (amplitude and phase) of the sample by iteratively combining low-resolution images illuminated from different angles in Fourier space. FP provides two advantages: 1- It provides high resolution images while maintaining a wide field of view (FOV), 2- It acquires quantitative phase information without using complex optical interference techniques. In this study, the aim was to detect liver fibrosis using Attention-based Deep Multiple Instance Learning (MIL) from phase images of liver biopsies obtained by FP microscopy. In this framework, very large liver biopsy phase images are divided into patches and fed to the model, and the model infers healthy or diseased states at the biopsy image level. Also, the attention layer used in this architecture ensures that different attention weights are assigned to each patch in each biopsy image. Thus, besides classification at the biopsy image level, the most significant patches on the slide leading to the positive label can be found. Thus, it can provide support to the pathologist in the detection of disease-related tissue regions. One of the aims of this study is to try to eliminate the subjectivity and inconsistency of traditional histopathological examination due to staining using QPI. Another aim of the study is to train a deep learning model to accurately diagnose liver fibrosis through the MIL framework with only biopsy-level labeling without any local labeling by the pathologist on phase images of liver biopsies. In addition, the attention mechanism ensures that areas of high diagnostic value are found. Thus, this approach has the potential to overcome some of the limitations of AI in medical diagnosis, such as interpretability and explainability. Phase images of liver biopsies with FP were obtained with high contrast and high resolution. As a result of deep learning studies on these data, an average of 88% accuracy was achieved and the most important patches were detected by the deep learning model. It was confirmed by the pathologist that these regions were associated with fibrosis. The results of this study show that FP and Attention-based MIL are very promising for digital pathology.