Artificial intelligence (AI) is rapidly advancing digital oncology.
A consensus report by 24 experts with first-hand experience in computational pathology/pathology AI (CPath/AI) reported that AI will improve diagnostic accuracy and that the daily tasks of pathology technicians will change significantly. By 2030, AI will be routinely and effectively used in pathology laboratories.
Two separate studies, recently published in eBiomedicine and The Lancet Digital Health, looked at: the prognosis of deep-learning-based scoring systems for tumor-infiltrating lymphocytes (TILs, which can be used as drug targets for the treatment of cancer) at different times of melanoma, a type of skin cancer; and the value of AI as a standalone reader for mammogram workflows.
Two studies on different cancers
The first study was done by researchers from dermatology at the University of Tuebingen in Germany, dermatology at the University of Heidelberg in Germany, and the Department of Pathology at Yale University School of Medicine in the United States. In the study, the researchers analyzed 321 primary melanoma and 191 metastatic samples using the deep learning algorithm NN192, an algorithm developed for the standard and digitized TILs scoring system "eTILs".
The researchers found that melanoma patients with low eTILs had more than twice the risk of distant metastases of cancer tissue than patients with high eTILs, and eTILs scores were reduced between primary melanoma and metastatic samples. Patients with eTILs scores ≤ 12.2% and were treated with anti-PD-1 immunotherapy had poor survival outcomes. This demonstrates that eTILs have predictive effects on primary melanoma samples, and that eTILs can predict response and survival outcomes in patients treated with PD-1.
In this regard, Roberto Salgado, co-chair of the International Immuno-Oncology Biomarker Working Group, said that accurate quantification of immune cells involves prognostic and predictive information, which is important for clinical pathways and customized treatment plans. In addition, computer evaluation results are much more accurate than manual evaluations.
The second study was done by Karin Dembrower and his team at the Department of Oncology Pathology at the Karolinska Institute in Sweden and Capio Sankt Göran Hospital in Sweden.
In this study, the research team included 55,581 women aged 40-74 years without breast implants based on regular breast cancer screening at Capio Sankt Göran Hospital from April 1, 2021 to June 9, 2022. The study followed the Swedish National Guidelines for mammography screening, in which two radiologists independently evaluated each participant's mammogram images and discussed by consensus in the event of any abnormal reading of either person to decide whether to proceed with further imaging tests. If cancer is still suspected on further testing, a biopsy sample is obtained, which is analyzed by a pathologist and definitively diagnosed.
In the study, while two radiologists read the films, the Insight MMG, an AI system, ran in the background as a standalone reader. Prior to the consensus discussion, radiologists did not have access to the Insight MMG for information, and in the consensus discussion, radiologists had access to information on all cases of the Insight MMG, including any local image findings, graphical contours, and corresponding AI anomaly scores.
The research team conducted four reading strategies, which looked at the actual diagnosis results of two radiologists with double reading (standard condition), one radiologist and the AI system double reading, the AI system single reading, and two radiologists and the AI system reading three readings. The results showed that compared to the standard situation, the cancer detection rate of a double reading of a radiologist and the AI system increased by 4% and the recall rate decreased by 4%; There was no significant difference in the detection rate of cancer in a single reading of the AI system, and the recall rate was reduced by 47%; Two radiologists and the AI system read the third reading slightly improved the cancer detection rate, improved the recall rate by 5%, and discussed consensus by nearly 50%.
The research team said that AI systems and humans will perceive certain different image features as suspicious cancer when reading films, so humans and AI systems work synergistically to improve the detection rate of breast cancer in mammograms. The AI system's single reading minimizes the psychological burden on participants due to multiple examinations, but this means that a large percentage of mammograms are never evaluated by doctors. Two radiologists and an AI system can detect cancer to the greatest extent, but this must be balanced with the increased cost of testing and a shortage of radiologists.
The market still needs to evolve
Roberto Salgado said digital biomarker testing can help clinicians make informed and personalized decisions in cancer treatment. However, as of 2023, few such products are mature and utilized on a large scale in the market.
On September 7, local time, Paige.AI, a US cancer diagnostic technology developer, announced a cooperation with the US technology company Microsoft (Microsoft) to build the world's largest image-based AI model and apply it to the development of digital pathology and oncology.
Coincidentally, on September 11, local time, US technology company Dell and the University of Limerick Digital Cancer Research Center in Ireland jointly developed AI platforms and digital twin technologies to promote the prediction and diagnosis of B-cell lymphoma.
"This is a very exciting start, and we look forward to the digital support from Dell's technical team to accelerate this project." Paul Murray, Professor of Molecular Pathology at the University of Limerick and Scientific Director of the Digital Pathology Division at the Digital Cancer Research Centre, said: "By working with Dell's technical team, we will be able to further our understanding of how cells go wrong during cancer development and find new ways to diagnose and treat cancer patients." ”