Highlights

[CVPR 2025, MICAD 2024] Rethinking Decoder Design: Improving Biomarker Segmentation Using Depth to Space Restoration and Residual Linear Attention 

We have presented complementary advances in decoder design for biomarker segmentation, with two papers accepted in highly regarded venues. One paper, “Rethinking Decoder Design: Improving Biomarker Segmentation Using Depth to Space Restoration and Residual Linear Attention,” was accepted at CVPR 2025, one of the most prestigious conferences in computer vision. The other, “Rethinking the Nested U Net Approach: Enhancing Biomarker Segmentation with Attention Mechanisms and Multiscale Feature Fusion,” was published as a MICAD 2024 book chapter under Springer’s LNEE series in 2025.

Our CVPR work introduces a decoder that combines depth-to-space upsampling with residual linear attention, allowing for finer spatial detail recovery and more robust feature alignment. In parallel, our MICAD chapter expands upon the nested U-Net framework by incorporating attention mechanisms and multiscale feature fusion, improving contextual awareness across image regions. Taken together, these approaches deliver accuracy improvements of 2.8% to 4.0% over strong baselines on diverse benchmark datasets including MoNuSeg, DSB Electron Microscopy, and TNBC.

Accurate biomarker segmentation, particularly of cell nuclei, is foundational in digital pathology and computational microscopy, as emphasized by the U.S. National Institute of Biomedical Imaging and Bioengineering. More reliable segmentation directly supports diagnostic accuracy, which is crucial for early cancer detection. The World Health Organization reports that cancer caused nearly 10 million deaths in 2020, underscoring the urgency of tools that can enable earlier and more precise intervention. By improving segmentation performance, our decoder designs help reduce suffering, lower treatment costs, and bring the benefits of computational pathology closer to clinical deployment.

Beyond diagnostics, our models have strong potential across multiple domains. In clinical pathology, they enable more accurate biomarker quantification and tumor grading, providing a basis for precision medicine and companion diagnostics. In population health, they can support more efficient cancer screening, helping triage cases and prioritize high-risk patients. In biomedical research and drug discovery, they accelerate cellular and tissue-level analysis, making treatment effects easier to quantify. In clinical trials, where reliable biomarkers serve as essential indicators for regulatory decision-making, our architectures can enhance the robustness of large-scale trial datasets.

By reimagining how decoders reconstruct spatial details and focus attention on critical regions through innovative upsampling, multiscale fusion, and attention, we have made biomarker segmentation both more accurate and more interpretable, bringing us closer to real-world pathology deployment.