2.10. Artificial Intelligence and the Polytechnic University of Catalonia–BarcelonaTech (UPC)

It is now recognized that artificial intelligence represents a turning point for society that is at least as significant as the Industrial Revolution. The ICS aims to be an autonomous and primary player in the creation of artificial intelligence algorithms to avoid dependance on a commercial solution. Therefore, we wished to set up an artificial intelligence platform tailored to the needs of the ICS. This platform was developed with free software and needed to be modeled to accommodate other artificial intelligence projects in addition to DigiPatICS.

Currently, the ICS has signed a collaboration agreement with the Image Processing Group (GPI) of the Polytechnic University of Catalonia–BarcelonaTECH (UPC) (imatge.upc.edu) for the development of artificial intelligence tools and platforms within the healthcare field, specifically in the field of medical imaging. The group belongs to the Intelligent Data Science and Artificial Intelligence Research Center (IDEAI), which is a hub created at the UPC in 2017 for the development of artificial intelligence. GPI has extensive experience in image processing and the development of artificial intelligence algorithms with a long history in the healthcare field.

GPI develops computer vision and deep learning (DL) tools to tackle WSI analysis tasks in DigiPatICS for stains, such as H&E, HER2, KI67, RE, and RP, as well as other immunohistochemical stains. DL technology that relies on instance and semantic segmentation architectures has the potential to provide high-quality results when properly trained. GPI has also introduced strategies, such as:

Dataset annotation: As the training database was limited in the first stage, a more classical computer vision strategy relying on morphological algorithms and machine learning (ML) tools produced proposals easier to validate by annotators and generated the ground truth needed to avoid limiting the performances of DL approaches.

Integration of AI algorithms: Integration into specialist workflow was facilitated by combining the systematic processing of a large number of WSIs with on-demand assessment by pathologists for improving the systematic results or for obtaining specific quantifications:

Nightly batch processing and inference on the WSIs yielded raw results, such as segmentation confidences and classification probabilities, as well as potential segmentation masks;

The results were integrated into 3DHISTECH ClinicalViewer using a specific plug-in and offered to the pathologists upon request when examining the slides;

The pathologists could select or deselect regions in the WSI to visualize and quantify the results or to fine-tune inference results (classification and segmentation) using sliders;

Pathologists could also select specific areas for further online analysis on inference servers to be performed during the session with the viewer at their workstations.

Pathologists in the analysis loop: The strategy in Point 3 allowed not only flexible interaction for online and on-demand analysis, but also for recovering information about pathologists by selecting specific regions of interest and tuning inference results for reports. This information represents invaluable data and comments to further improve the annotated datasets.

Training in the ICS development servers: A specific committee formed by AI specialists and clinicians periodically reviewed the comments and data produced by pathologists using the viewer. This committee decided on the feedback and data and set strategies for incremental training and improvement of the DL network architecture involved with continuously improving the inference results.

The former strategies facilitated the usage of AI tools in the daily work of pathologists, as well as productivity.

As mentioned, all images were stored in a central repository, SIMDCAT, where they were available for AI training after dissociation. AI training was performed in this central repository, where a large number of whole slide images were readily at hand. GPI researchers did not have direct access to sensitive clinical data, and datasets and whole slide images always remained within the ICS infrastructure. In fact, it should be noted that the entire circuit in production moved through an isolated dedicated network in an intranet environment, and the training environment was in a separate VLAN and intranet infrastructure. Once the inference algorithms were trained, they were run on-premise at each hospital DPC. In no case was the information transferred to external DPCs.