A recent study introduced a cutting-edge AI-powered pathology platform that can help doctors diagnose and evaluate lung cancer in patients quickly and accurately. Developed by a team of researchers at the University of Cologne’s Faculty of Medicine and University Hospital Cologne, the tool provides fully automated and in-depth analysis of benign and cancerous tissues, for faster and more personalized treatment.
Lung cancer is known for high mortality rates, but precise diagnostics and personalized treatments lead to better outcomes for patients. Traditionally, oncologists manually examine tissue samples under a microscope to identify cellular and structural characteristics that reveal cancer. However, even with expert analysis, the process is time-consuming, subjective, and prone to variability, which can lead to misdiagnosis.
The researchers developed a deep-learning-based multi-class tissue segmentation platform that automatically analyzes digitized lung tissue samples. It screens for cancer and provides cellular details of the region.
The AI model was trained and validated on a large dataset from six institutions, totaling 4,097 annotated slides from 1,527 patients.
According to study senior author Yuri Tolkach, “The algorithm can differentiate between 11 tissue types, ranging from tumor tissue, tumor-associated classes (e.g., tumor stroma, necrotic debris, mucin) to cartilage and lymphatic tissue. It showed very high pixel-wise accuracy for segmentation of different classes with an average Dice Score 0.893.”
The researchers used the University of Cologne’s high-performance computing cluster equipped with 12 NVIDIA V100 GPUs, four NVIDIA A100 GPUs on the pathology institute’s AI server, and PC stations equipped with NVIDIA GeForce RTX 3090 and NVIDIA RTX 4090 GPUs.
The setup enables quick analysis of entire slide images. It takes about 1 to 5 minutes to analyze each whole-slide image ranging from 200 to 2000 Mb.
“The formation of our research group and our first large cancer study published in Nature Machine Intelligence was made possible through an NVIDIA Quadro P6000 GPU grant from the NVIDIA Academic Grant Program,” Tolkach said.
The AI tool can also reveal detailed characteristics of tumor and immune cells in the cellular environment. This unveils how the cancer is interacting within the body.
Identifying subtle patterns and relationships within the tissue sample not visible to the naked eye could help inform more precise and effective treatments, and offer insight into a patient’s response to a specific cancer therapy.
The code used in this study is available on GitHub.
Read the study Next-generation lung cancer pathology: Development and validation of diagnostic and prognostic algorithms.
