(B-52) Leveraging Transfer Learning for Dll4 Expression Assessment in Triple-Negative Breast Cancer: A dynamic enhanced near-infrared Fluorescence Imaging Approach with Time-Series Image Representations

Introduction:: Delta-like canonical notch ligand 4 (Dll4 ) is a transmembrane protein involved in the Notch signaling pathway, playing a critical role in vascular development and angiogenesis in tumors. Non-invasive detection of Dll4 expression levels is crucial for effective decision-making in cancer therapy, particularly anti-angiogenic therapies. This study aims to develop a novel framework for stratifying Dll4 expression levels in triple-negative breast cancer using dynamic enhanced near-infrared (DE-NIR) fluorescence imaging with indocyanine green (ICG) and transfer learning on time-series image representations.


Materials and Methods:: We generated two rat strains, Salt sensitive (SS(n=8)) with high systemic Dll4 expression and consomic SSBN3(n=18) with third chromosome substitution from Brown Norway with low Dll4 expression. We also constructed two novel SSBN3 congenic xenograft host strains (CG1(n=12), CG2(n=12)) by introducing segments of BN chromosome 3 into the genetic background of the parental SS strain, such that CG1 inherited Dll4 locus from SS and CG2 from BN. A total of 50 animals bearing MDA-MB-231 triple-negative breast cancer xenografts were imaged using a whole-body imaging platform for rats. Animals were injected intravenously via tail vein with ICG (0.75mg/kg, MP Biomedicals) and simultaneously imaged using a PIMAX4 ICCD camera (Princeton Instruments), equipped with a 25mm lens (Navitar, DO-25, 0.95 f-stop), with a holographic notch rejection filter and a 830 nm Long-pass filter (ThorLabs), excited by a 785nm diode laser (~5mW/cm2), at a frame rate of 10.6 fps for 6 minutes. Respiratory motion correction was performed with Fourier-Wavelet methods, and tumor ROIs were automatically drawn on principle component decomposed images. Time-series data were converted into images using the Gramian Angular Field (GAF) representation and Markov Transition Field (MTF) method. We applied transfer learning using convolutional neural networks (CNNs) for classification of Dll4 expression levels on these image representations, training the model on consomic rats (SS and SSBN3) and testing it on congenic rats (CG1 vs CG2) to ensure its generalizability.


Results, Conclusions, and Discussions:: The transfer learning approach on image representations derived from time-series data exhibited outstanding performance in stratifying Dll4 expression levels in tumors. Our model achieved a sensitivity of 94% and a specificity of 92% in identifying Dll4 expression alterations. Additionally, the area under the receiver operating characteristic (ROC) curve was 0.97, highlighting the robustness and discriminative power of the proposed framework. These results suggest that the combination of DE-NIR imaging, image representation techniques (GAF and MTF), and transfer learning using convolutional neural networks can accurately and noninvasively assess Dll4 expression levels in tumors, enabling effective decision-making for personalized cancer therapy.

DE-NIR imaging presents an effective alternative for characterizing germline-dependent vascular phenotypes compared to dynamic vascular imaging techniques such as DCE-MRI and perfusion CT, which have limitations like low temporal resolution and the need for specialized hardware and software. By converting time-series data into image representations and leveraging transfer learning, our approach can potentially stratify patients for Dll4 targeted therapies, aiding in effective decision-making for cancer therapy. Further validation with cross-sectional dynamic contrast-enhanced imaging and other time-series to image conversion methods is required to extend this approach to clinical modalities.


Acknowledgements (Optional): :

References (Optional): :