Johns Hopkins University Reconstructs 3D Tissue Architecture to Reveal Hidden Cancer Precursors

At Johns Hopkins University, a multidisciplinary team of engineers, cancer biologists, and pathologists sought to address a major limitation in cancer research: the inability to visualize and analyze microscopic precancerous lesions in three dimensions. Pancreatic intraepithelial neoplasia (PanIN), a common precursor to pancreatic cancer, is too small to be detected by conventional imaging techniques and traditionally studied using 2D histological slides. This approach restricted researchers’ understanding of the true structure, frequency, and biological significance of these lesions.

To overcome these challenges, the team developed CODA, a computational platform that reconstructs 3D tissue models from serial histological sections. The process required solving complex image alignment challenges, as the physical sectioning of tissue introduces artifacts such as tearing, folding, and distortion. CODA uses Image Processing Toolbox™ and Deep Learning Toolbox™ to implement nonlinear image registration, nuclear coordinate detection, and deep learning–based segmentation using the DeepLabv3+ algorithm with a ResNet-50 backbone. The workflow involves downsampling images, manually annotating training data, and applying deep learning to automatically segment anatomical structures. Transformation matrices calculated on lower resolution images are used to register high-resolution segmented images, enabling accurate 3D reconstructions.

With CODA, the team generated anatomical maps of human pancreas tissue, revealing that PanIN lesions are more prevalent and morphologically complex than previously recognized. CODA also enabled integration of histological data with genomic, proteomic, and transcriptomic profiling, providing a comprehensive view of tissue architecture and molecular characteristics. Through integration of 3D morphology and genomics, the team found that PanINs develop independently from one another, with each anatomically separate PanIN possessing a different mutation in the oncogene KRAS. The platform’s flexibility has allowed its application to other organs and research areas, including mapping the developing kidney, fallopian tubes, and heart, and designing biomimetic organoid models.

Future plans include expanding CODA’s capabilities with nuclear segmentation, slide interpolation, virtual staining, and user-friendly app development to further support cancer and developmental biology research.