(B-78) Comparative Analysis of Machine Learning & Transfer Learning for Personalized Therapy: Assessing Dll4 Heterogeneity in Triple-Negative Breast Cancer via Dynamic Enhanced Near-Infrared Imaging

Introduction:: Dll4, a transmembrane protein involved in the Notch signaling pathway, plays a critical role in vascular development, angiogenesis, and cancer therapy efficacy. Non-invasive detection of Dll4 expression in tumors is essential for personalized cancer therapy and effective decision-making. This study aims to develop a robust and generalizable framework for Dll4 expression stratification in triple-negative breast cancer using Dynamic Enhanced Near-Infrared (DE-NIR) fluorescence imaging with indocyanine green (ICG). We compare two different approaches: machine learning on 1D time series and transfer learning on time series converted to images using the Markov Transition Matrix, highlighting their potential in facilitating personalized cancer therapy and effective decision-making.


Materials and Methods:: We generated two rat strains, Salt sensitive (SS; n=8) with high systemic Dll4 expression and consomic SSBN3 (n=18) with a third chromosome substitution from Brown Norway, resulting in low Dll4 expression. Additionally, two novel SSBN3 congenic xenograft host strains (CG1; n=12, CG2; n=12) were constructed by introducing segments of BN chromosome 3 into the genetic background of the parental SS strain, such that CG1 inherited the 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 intravenously injected with ICG via the tail vein and simultaneously imaged using a PIMAX4 ICCD camera, equipped with a 25mm lens, a holographic notch rejection filter, and an 830 nm Long-pass filter, excited by a 785nm diode laser at a frame rate of 10.6 fps for 6 minutes. Respiratory motion correction was performed.



For both machine learning and transfer learning approaches, principal component decomposed images along the time axis and modified PCA techniques were employed to visualize, segment, and analyze tumors and normal ROIs. In the machine learning approach, discriminative features were generated from the ROIs for classification, and algorithms such as decision trees, support vector machines, and neural networks were applied to the generated features for classification. For the transfer learning approach, computer vision techniques like GAF and MTF were used to encode time-series data as images, preserving temporal dependencies, and enabling the use of visual algorithms like CNN and ResNet for classification.


Results, Conclusions, and Discussions:: Both the machine learning and transfer learning approaches accurately identified host Dll4 expression alterations with sensitivity and specificity above 90%, demonstrating their potential for noninvasively assessing Dll4 expression levels in tumors for personalized cancer therapy. By leveraging advanced computer vision techniques, the transfer learning approach showed promising results in capturing granularity within the time-series data and enhancing prediction performance.



This comparative analysis of machine learning and transfer learning approaches highlights their potential in facilitating personalized cancer therapy and effective decision-making for triple-negative breast cancer patients. By noninvasively stratifying Dll4 expression levels, clinicians can better tailor treatment strategies, monitor therapeutic responses, and ultimately improve patient outcomes. Furthermore, the developed framework can be extended to clinical modalities upon validation with cross-sectional dynamic contrast-enhanced imaging.


Acknowledgements (Optional): :

References (Optional): :