Researchers develop lightweight AI system to improve childhood tumor diagnosis from pathology images

Scientists have developed a new computational approach to improve the diagnosis of peripheral neuroblastic tumors (pNTs), aggressive childhood cancers that require precise classification to guide treatment decisions. The system, called CoPath, combines neural network innovation with domain knowledge from pathology to analyze whole-slide histopathology images—the microscopic tissue samples that pathologists examine daily.

The Challenge#

Accurate diagnosis of neuroblastic tumors is critical for determining prognosis and selecting appropriate therapies for children. Yet pathologists face several obstacles when categorizing these tumors under the microscope. The tissue appearance varies dramatically within a single tumor and between patients, few reference samples exist from pediatric cohorts, and different observers may reach different conclusions examining the same slide. Additionally, existing computational tools that process these large-format digital slides demand substantial computing power.

How CoPath Works#

The research team's solution centers on two components working in tandem. CoHisNet, the first part, is a compact neural network designed to analyze small tissue patches extracted from whole-slide images. Rather than relying solely on traditional neural network building blocks, the system incorporates Kolmogorov-Arnold Networks—a newer mathematical approach that better captures complex, nonlinear relationships in tissue data. The network is trained to examine tissue structure at multiple scales simultaneously, capturing both the broader architectural patterns and the fine details of individual cells.

PathVote, the second component, synthesizes these patch-level classifications into a final diagnosis for the entire slide. Crucially, it does so by incorporating knowledge about how normal tissue components are typically organized in healthy versus diseased samples—essentially encoding expert pathology reasoning into the decision-making process.

Performance and Efficiency#

Testing on a private cohort of pediatric neuroblastic tumors and the publicly available BreakHis breast cancer dataset showed that CoPath matched or exceeded the accuracy of existing approaches. Importantly, it did so while maintaining substantially lower computational requirements than comparable systems, including recent large-scale foundation models fine-tuned for pathology work. The researchers have made their code publicly available to support further development and validation by the research community.