Pathologists base their routine clinical practice on the recognition, semi-quantification and integration of morphological patterns according to predefined criteria dictated by the clinical context. These data are then classified and summarized in the histopathological report. The natural differences in visual perception, data integration and judgement between each pathologist makes Pathology a subjective discipline. The first efforts to extract objective measures from microscopic images were produced in cytology. The introduction of scanners in the 1990s allowed the creation of digitized images of tissue slides (Whole Slide Images, WSI) and led to a growing interest in the application of Artificial Intelligence (AI) in Pathology, which previously happened in the field of Radiology [1,2,3]. Machine Learning (ML) was the first AI technique applied in the field of Pathology. ML-algorithms were based on engineering specific morphological features used by pathologists for a specific task, for example cell size, nuclear shape and cytoplasm texture to discriminate between benign and malignant tumors. The main limitation of ML was that annotation of specific features was timeconsuming and sometimes anchored to the original problem domain (not scalable to other problems). The introduction of Deep Learning (DL) boosted the explosion of AI applied to histopathology. Indeed DL approaches can learn directly from raw data (WSI) and do not rely on the effort to engineer a feature anchored to the problem as in ML. Nonetheless, even if DL does not need pre-existing assumptions, raw data might need a certain degree of control. DL approaches can be further subdivided into strongly supervised, weakly supervised or unsupervised. In strongly supervised DL, several patches extracted from WSI should be labelled by pathologist with the class that the model is intended to predict. In weakly supervised approaches annotations are made at image level (each WSI is labelled with a specific class). In these models, a single label of interest might be used to develop an algorithm: the presence or absence of tumor (the most common); the presence of a somatic mutation; the clinical outcome. While easier to be labelled, series evaluated by weakly supervised DL models usually need to be larger than those used for strictly supervised approaches. As regard to the dimension of the dataset it is interesting to observe that convolutional neural network (CNN) can be trained on a source task and then be reused on a different target task. This technique, known as transfer learning, can be extremely useful when the data for the target task is scarce but a larger dataset is available to train the source task [4,5]. Finally, in unsupervised DL, the learning examples are provided with no associated labels. This approach, which to date has been applied in a very limited number of examples [6] is an active area of machine learning research.

The term Digital Pathology (DP) encompasses all the digital technologies related to the introduction of WSI that allow improvements and innovations in the workflow of a Pathology Department [7]. AI-based tools are part of these technologies: starting from WSI they promise to improve pathologists’ activity to recognize, quantify and integrate information written in the tissue (tissue biomarker) for diagnostic, prognostic and predictive purposes. In order to develop such AI-based tools, researchers need specific software to work on WSI. These include color normalization, focus quality assessment, standard tiles extraction, object classification, region segmentation and counting. Table 1 lists some among the most popular software used to these aims. A list far to be complete taking into consideration that the open source Github (https://github.com/, accessed on 26 February 2021) returns more than 100 software solutions tagged by “digital pathology” while a recent forecast for DP market (https://www.marketsandmarkets.com/Market-Reports/digital-pathology-market-844.html#:~:text=%5B210%20Pages%20Report%5D%20The%20global,13.8%25%20during%20the%20forecast%20period.&text=However%2C%20a%20lack%20of%20trained,growth%20in%20the%20coming%20years, accessed on 26 February 2021) catalogue more than 20 companies working on this topic. The platform BIII (https://biii.eu/, accessed on 26 February 2021) developed by the Network of European Bio-image Analyst NEUBIAS proposes to help the researcher to compare this plethora of solutions focusing on the problems the tools can solve rather than comparing their technical aspects. Finally, it should be remembered that AI-tools which are to be implemented in clinical practice are subject to strict regulation. The new Conformite Europeénne—in vitro diagnostic device regulation (CE-IVDR) will significantly impact this process in EU from 2022. Table 2 lists CE approved AI-based tools and their application field.

In the setting of Personalized Medicine (PM) a biomarker is every piece of information used to recognize a specific subset of a larger population with diagnostic, prognostic and/or predictive purposes. Tissue biomarkers are information written in the tissue, which can be interpreted by properly only by Pathology. Histotype, Grade and Stage of malignant tumors are “classic” tissue biomarkers: recognizing patients affected by early stage disease and with specific favorable histotype, represent the origin of PM. Estrogen and progesterone receptors, Ki67 and HER2/neu, aka the biological profile of breast cancer, were the first examples, in the 2000s, of “new” tissue biomarkers: the expression of these phenotypical indicators allowed to select properly medical treatment and predict outcome of patients affected by breast cancer. In the last two decades, several other biomarkers have been discovered at different deepness (cellular, subcellular, molecular) in distinctive cancer population (neoplastic cells, cancer-associated immune cells, etc.). The increasing number of tissue biomarkers and the complexity of their evaluations, strongly encourages the use of AI based tools in the process of evaluation, to reinforce the role of Pathology in PM. [8]. Moreover, AI algorithms have been used to discover information which are ignored by human review of an H/E image and use them to make specific prediction as shown in Figure 1. These AI-based biomarkers include the prediction of treatment response [9,10], somatic mutations [11], or patient survival [12,13]. In the following sections, we will present how AI has been used to support tissue biomarkers evaluation in specific field of Pathology; also, we will give an insight to the intriguing field of AI-based biomarkers.

The application of AI in the real clinical setting is still limited by several issues.

The low level of digitization likely represent the first critical issue. A recent survey in England revealed that an access to a complete DP workstation was available in less than 30% of Institutions; the most common applications were teaching, research and quality assurance while direct clinical use was less widespread with consultations outdating primary diagnosis [60]. Lack of robustness and general applicability is another restriction to the application of AI in the daily practice. Most of the available AI models have been trained on small data sets and can present a 20% drop of performance when applied in a setting different from where they had been originated. Dataset can be enlarged using specific technical solution such as the transfer learning approach [4,5]. Another possibility will be the development of open sources datasets such as those already hosted by the Cancer Genome Atlas, the Cancer Imaging Archives and Grand Challenges. Recently a call for a “central repository to support the development of artificial intelligence tools”, was proposed by the H2020 program IMI2-2019-18. This dataset aims to endorse WSI, molecular and clinical data and to serve as raw data for the scientific community. In addition to large series, another prerequisite to AI elaboration is well-annotated ground truth generated by expert pathologists with specific, time-consuming sessions. This is in striking contrast to one of the popular motivations for the introduction of AI in pathology, namely the shortage of pathologists. Technical solutions, such as the data augmentation [61], image synthesis [62] and the adoption of weakly supervised or unsupervised DL model, have been suggested or are actively explored to fix this paradox. Finally variations in staining procedures, tissue types and scanners might be relevant to obtain higher performing AI systems [63]. Low adherence amongst the use of the AI models by pathologists can be another source of limitation. In their clinical activity, skilled pathologists examine the slide in two steps: first a scanning magnification to understand the general context of the disease; then a more careful evaluation with progressively higher magnifications to prove their general impression. Most AI models are developed using smaller tiles, rather than entire WSI, as input data, missing the efficacy warranted by the dual approach of the pathologist. To avoid this drawback recent studies have suggested the introduction of networks trained with images obtained at different magnification [64,65]. A different technical solution was proposed by Lin et al. [66] who introduced in the neural network a further layer aimed to reconstruct the loss occurred in max pooling layers. Also the interpretation of AI decisions, sometimes referred to as the ‘black box’ problem [67], is a relevant concern to complete adoption of AI. To overcome this, it has been suggested to link the solution proposed by AI models to different type of visual maps describing the abundance and morphology of features (necrosis, pleomorphism, etc.) known by pathologists [68,69,70].

On the other hand, a full integration of AI in Pathology is likely to represent a milestone of digital health. The introduction of AI as a device assisting pathological diagnosis is expected to reduce the workload of pathologists; to help standardize the otherwise subjective diagnosis that can lead to suboptimal treatment of patients; to help discovery new perspectives in human biology, and progress on personalized diagnostics and patient care [71]. In the current state AI platforms are developed with different functions, requiring users to launch different software for each purpose or to repeatedly download and upload images. The development of a simplified user interface, either on the WSI viewer or on the LIS (laboratory integration system), is a key factor to the successful implementation of AI at clinical level. Moreover, the availability of platforms integrating different software solutions with multiple clinical data to suggest prognosis and/or the choice of therapy will be another substantial benefit of digitization and provide an additional sanity check on AI generated predictions. Another relevant aspect of progressive digitalization and introduction of AI in Pathology will be a more integrated approach with radiomic. The latter rests on the hypothesis that data derived from digital radiological images have a correlation with the underlying biological processes and that this correlation can be caught by AI better than the visual interpretation. Pioneering studies recently explored the associations between radiomic and its counterpart on digital Pathology images (i.e., pathomic) in lung and breast cancer and revealed promising preliminary results [72,73].

When such advanced (“next generation”) pathological diagnosis enter medical practice, it is likely that the demands of clinicians would not be satisfied with the level of current pathological diagnosis offered by pathologists using solely a microscope. Pathologists who reject digital Pathology and AI may face a diminished role in the future of Pathology practice.