AI tool optimizes disease diagnosis by analyzing biopsies

The combination of artificial intelligence and human collaboration is transforming the analysis of tissue (anatomopathological) samples obtained from biopsies or surgical specimens. A study published in the journal Nature Communications, with participation from the D’Or Institute for Research and Education (IDOR), presented an innovative collaborative platform capable of segmenting digital tissue images with high precision—even with limited manual data—and identifying markers related to the prognosis of patients with melanoma, a type of skin cancer. Dr. Lidiane Vieira Marins, a pathologist at Rede D'Or and an IDOR researcher, was among the authors representing Brazil in the study.

Histological analysis of tumor tissue is a crucial step in the diagnosis and treatment of various types of cancer. This process relies on the interpretation of images by pathologists who manually examine slides stained with hematoxylin and eosin (H&E). With the ability to digitize these slides and the advancement of AI, new possibilities are emerging to automate and standardize this analysis. However, training robust algorithms requires enormous volumes of data annotated by specialists, a process that is both costly and time-consuming.

To overcome this limitation, researchers developed PHARAOH (PHenotyping and Regional Analysis Of Histology), a free, interactive platform for the computational analysis of histological tissues. The tool uses Deep Learning models based on less detailed annotations from a small number of cases, which are sufficient to extract relevant clinical and biological patterns on a large scale.

The study demonstrated that with just seven initial samples of cutaneous melanoma, PHARAOH was able to identify and classify tumor and adjacent regions with high precision, creating a database of over 23,000 automatically annotated images. From these images, the researchers trained a convolutional neural network (CNN) to segment tumor lesions in other patients. This type of algorithm is designed for pattern recognition, making it ideal for image processing and analysis.

The platform's key advantage lies in its combination of automation with targeted human intervention. The process begins by dividing the complete images into small fragments called "tiles." These tiles are smaller pieces of the histological image (e.g., from a tissue slide), which simplifies the analysis.

Next, the system groups these fragments based on morphological similarities. In other words, fragments that "look alike" in terms of color, shape, texture, or cellular patterns are automatically placed together by the system using unsupervised clustering algorithms.

Instead of having to label each fragment individually, the specialist analyzes these pre-formed groups and provides a single label for the entire group. These human-evaluated definitions are then used to train a personalized machine learning model, which learns to automatically recognize the patterns associated with each label based on the provided samples. This model can be continuously refined by specialists.

"While melanoma was used as the initial example, the platform has been successfully applied to other types of cancer, demonstrating its potential for widespread use in digital pathology. In addition to supporting diagnosis, the platform aids in predicting clinical outcomes and identifying morphologically relevant biological patterns," adds author Dr. Lidiane Vieira Marins.

According to the article, PHARAOH was designed as a collaborative encyclopedia for computational pathology models. Any researcher can submit images, train models, and share the results with the scientific community. The platform already offers integrated features for tissue segmentation, cellular phenotyping, and the export of regions of interest for analysis with external tools.

Testing the Tool for the Study

To test the clinical relevance of PHARAOH, researchers calculated a tumor-infiltrating lymphocytes (TILs) score—immune cells present at the tumor margins—in 385 samples available for analysis in the database. A higher presence of these cells was associated with a significantly longer median survival time (109 months) compared to patients with less infiltration (66 months). This finding reinforces the prognostic value of TILs and shows that the platform can capture biologically relevant signals from images.

The system showed performance similar to manual reading methods using immunohistochemical markers, with high concordance rates in tissue segmentation. This indicates that the platform can be a reliable alternative for accelerating histological research without compromising accuracy.