Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Biomedical Imaging and Instrumentation

Biomedical Imaging and Instrumentation - Poster Session C & D

U39 - AI-Based Quantification of Nuclear Heterogeneity in Normal and Cancerous Breast Cells

Friday, October 10, 2025

10:00 AM - 11:00 AM PDT

Location: Exhibit Hall F, G & H

Presenting Author(s)

Yujin Sung

Graduate Student
KAIST, United States

Co-Author(s)

Seok Hoon Ham, ME

P.h.D candidate in Mechanical Engineering(KAIST)
Mechanical Engineering Department, Korea Advanced Institute of Science and Technology
Yuseong-gu, Daejeon, Republic of Korea

Primary Investigator(s)

Jennifer H. Shin

Professor
KAIST
Daejeon, Republic of Korea

Introduction: : The nucleus reflects intrinsic cellular states such as genomic stability, chromatin architecture, and transcriptional activity. Although nuclear shape is a key factor in cancer diagnosis, it remains challenging to quantify the extent of variation between cells on a large scale (Abel et al., 2024). Traditional morphological descriptors (e.g., area, circularity) are often insufficient to capture the complex nuclear heterogeneity associated with cancer (Radhakrishnan et al., 2017). To overcome this limitation, this study adopts a deep learning approach capable of autonomously discovering subtle morphological patterns beyond human-defined parameters. Breast epithelial cells (MCF10A and MCF7) were selected as biologically meaningful models because nuclear pleomorphism, a hallmark of breast cancer, is a clinically significant criterion explicitly incorporated into histological grading systems such as the Nottingham grade (Mercan et al., 2022). An explainable AI (XAI) framework was further applied to extract high-dimensional morphological features directly from nuclear images, enabling quantitative comparison and interpretation of phenotypic differences between normal and cancer-derived nuclear architectures.

Materials and

Methods: : MCF10A (non-tumorigenic) and MCF7 (cancerous) human breast epithelial cells were cultured on 40 mm plasma-treated confocal dishes. After coating with Matrigel (1:49 dilution in serum-free DMEM) and 1-hour incubation at room temperature, the solution was aspirated to leave a thin residual layer simulating a soft ECM-like environment. Cells were seeded at 1×10⁵ (MCF10A) and 1.5×10⁵ (MCF7) cells per dish to achieve ~60–80% confluency, forming moderate islands rather than a compact monolayer. Following 1–2 days of incubation at 37°C and 5% CO₂, nuclear morphology was assessed via DAPI staining. A total of 2,526 MCF10A and 2,606 MCF7 nuclei were imaged using fluorescence microscopy, capturing both DAPI and corresponding 100× bright-field images (Fig. 1a).

Nuclear images were processed through a ResNet34-based supervised learning model trained for cell type classification. To explain black-box predictions and quantify nuclear heterogeneity, XAI techniques were applied. Integrated Gradients (IG) attribution analysis, combined with a multi-baseline strategy, was employed to identify discriminative image regions. Attribution maps were aggregated over superpixels segmented via the Simple Linear Iterative Clustering (SLIC) algorithm, and only high-attribution regions filtered by Otsu thresholding were retained. Mean IG scores were assigned to selected superpixels, allowing for localized interpretation of the morphology contributing to model predictions. High-dimensional features from the model’s embedding layer were projected into principal component space to quantify heterogeneity in nuclear morphology.

Results, Conclusions, and Discussions:: The ResNet34 model trained on nuclear morphology images achieved a test accuracy of 99.8% across 411 samples, demonstrating strong generalization performance and suitability for downstream explainability analysis (Fig. 1b, c). This high performance indicates that the model effectively captured biologically meaningful representations of nuclear structure. Model interpretability was achieved using XAI methods applied to global average pooling (GAP) vectors, which were subsequently embedded into principal component (PC) space to explore nuclear morphological heterogeneity. As illustrated in Fig. 2a, superpixels with high attribution scores were selectively occluded to generate these vectors. Embeddings originating from the same nucleus formed compact clusters (Fig. 2b), indicating robust internal consistency and preservation of single-cell identity. Additionally, MCF10A and MCF7 nuclei exhibited distinct separation along the PC1 axis, suggesting effective model capture of cell-type–specific morphological features.

To quantify nuclear heterogeneity, convex hulls were calculated around GAP vector clusters for each cell type (Fig. 2c). MCF7 cells exhibited a significantly larger convex hull area and greater dispersion compared to MCF10A, indicating elevated heterogeneity associated with malignant phenotypes. The observed morphological differences likely result from fundamental alterations in nuclear architecture driven by malignant transformation. Specifically, MCF7 cells displayed increased nuclear irregularity, possibly due to chromatin remodeling or genomic instability–common characteristics of cancer cells. Further analysis is necessary to elucidate whether these morphological features correlate directly with known biological phenotypes or molecular markers.

The increased nuclear heterogeneity observed in MCF7 cells likely reflects critical cancer phenotypes, such as differential drug responses, therapy-induced senescence, or increased cellular plasticity, reflecting functional consequences of nuclear remodeling in malignancy. By converting black-box model predictions into biologically interpretable features, the framework bridges computational phenotype analysis with mechanistic insights into cellular function. This approach thus contributes toward developing scalable and mechanistically informative single-cell profiling tools. Future work will focus on mapping these AI-derived morphological features to known molecular phenotypes such as drug response variability, metastatic potential, epigenetic states, and dynamic cellular processes, including epithelial-mesenchymal transition (EMT) or therapy-induced senescence. Ultimately, this method provides a versatile platform with significant potential applications in disease characterization, personalized treatment stratification, and broader translational oncology research.

Acknowledgements and/or References (Optional): : Acknowledgement
This research was supported by the Boston-Korea Collaborative Research Program funded by the Ministry of Science and ICT and the Ministry of Health and Welfare of Korea (Project Number: RS-2024-00468873).

References
Abel, J., Jain, S., Rajan, D., Padigela, H., Leidal, K., Prakash, A., Conway, J., Nercessian, M., Kirkup, C., Javed, S. A., Biju, R., Harguindeguy, N., Shenker, D., Indorf, N., Sanghavi, D., Egger, R., Trotter, B., Gerardin, Y., Brosnan-Cashman, J. A., . . . Taylor-Weiner, A. (2024). AI powered quantification of nuclear morphology in cancers enables prediction of genome instability and prognosis. Npj Precision Oncology, 8(1). https://doi.org/10.1038/s41698-024-00623-9

Mercan, C., Balkenhol, M., Salgado, R., Sherman, M., Vielh, P., Vreuls, W., Polónia, A., Horlings, H. M., Weichert, W., Carter, J. M., Bult, P., Christgen, M., Denkert, C., Van De Vijver, K., Bokhorst, J., Van Der Laak, J., & Ciompi, F. (2022). Deep learning for fully-automated nuclear pleomorphism scoring in breast cancer. Npj Breast Cancer, 8(1). https://doi.org/10.1038/s41523-022-00488-w

Radhakrishnan, A., Damodaran, K., Soylemezoglu, A. C., Uhler, C., & Shivashankar, G. V. (2017). Machine learning for Nuclear Mechano-Morphometric biomarkers in cancer diagnosis. Scientific Reports, 7(1). https://doi.org/10.1038/s41598-017-17858-1