A team of researchers from the University of Maine has developed an innovative artificial intelligence (AI) system designed to improve the accuracy of breast cancer diagnoses from tissue samples. The system, known as the Context-Guided Segmentation Network (CGS-Net), aims to expedite the identification of cancerous cells, potentially preventing delays that can lead to adverse health outcomes.
Led by doctoral candidates Jeremy Juybari and Josh Hamilton, the research introduces a deep learning architecture that mimics the analytical approach of human pathologists. This advancement is particularly significant given that breast cancer is the second leading cause of cancer-related deaths among women, affecting approximately one in eight over their lifetimes. Despite advancements in diagnostics, the traditional method of microscopic examination of stained tissue samples remains time-consuming and dependent on expert interpretation.
In many regions of the world, particularly in developing countries like India, access to trained pathologists is severely limited. Currently, two-thirds of the world’s pathologists are concentrated in just ten countries. This disparity contributes to diagnostic delays that can lead to preventable deaths; about 70% of cancer-related deaths in India are associated with treatable conditions, exacerbated by the lack of timely diagnostic services.
Innovative Technology Behind CGS-Net
The CGS-Net employs a dual-encoder model that replicates the way pathologists assess tissue slides. Traditionally, pathologists zoom in and out of images to gather necessary data. This model allows CGS-Net to analyze both high-resolution images that capture cellular details and lower-resolution images that provide broader tissue context simultaneously. Each image patch shares a central pixel, enabling precise alignment of the two views.
During testing, the researchers trained CGS-Net using 383 digitized whole-slide images of lymph node tissue to identify signs of breast cancer and differentiate between healthy and cancerous tissue. The results showed that CGS-Net consistently outperformed traditional single-input models, demonstrating its potential for more accurate diagnostics in clinical settings.
In his remarks, Juybari emphasized the importance of integrating local and contextual tissue information to enhance cancer prediction accuracy. “By introducing a unique training algorithm and an innovative initialization strategy, our research illustrates how incorporating surrounding tissue context can significantly bolster model performance,” he stated.
Future Prospects and Broader Implications
While the initial study focused primarily on binary cancer segmentation, the team envisions expanding the system’s capabilities. Future research may explore the integration of additional image resolutions and apply CGS-Net to multiclass tissue segmentation. There is also potential for the architecture to incorporate multimodal data, such as radiology scans or molecular profiles, which could further enhance diagnostic accuracy.
Beyond its technical innovations, this project highlights the interdisciplinary strengths of the University of Maine, combining expertise in engineering, computing, and biomedical science to address global health disparities. As cancer diagnosis increasingly shifts towards digital methods, tools like CGS-Net are poised to augment, rather than replace, the expertise of medical professionals.
By teaching machines to analyze tissue samples similarly to how doctors do, researchers hope to pave the way for improved access to early and accurate cancer detection. With the dataset and source code made publicly available, the initiative encourages collaboration and transparency within the scientific community.
For further details, refer to the paper published in Scientific Reports by Juybari and Hamilton, along with faculty researchers Andre Khalil, Yifeng Zhu, and Chaofan Chen. The study, titled “Context-guided segmentation for histopathologic cancer segmentation,” is set to be published in 2025 and can be accessed via DOI: 10.1038/s41598-025-86428-7.
