About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Nanotechnology for Energy, Environment, Healthcare and Industry
|
| Presentation Title |
Deep Learning on Ultrasound Enables Distinction of Immunocompetent and Immunodeficient Tumors |
| Author(s) |
Navin Manjooran, Peter Warren, Faraz Chimani, Ashish Ranjan |
| On-Site Speaker (Planned) |
Navin Manjooran |
| Abstract Scope |
Accurate assessment of tumor immune competence is essential for predicting therapeutic response, yet current approaches rely on invasive sampling. Here, we present an ultrasound-based deep learning framework that integrates segmentation and classification models to non-invasively distinguish immunocompetent from immunodeficient tumors in vivo. Using C57BL/6 (immunocompetent) and RAG knockout (immunodeficient) murine models, we collected 30–60 B-mode ultrasound images per tumor, yielding a dataset of 52 tumors (21 RAG KO, 31 C57). A convolutional neural network (CNN) achieved a Dice coefficient of 85% for automated tumor segmentation relative to manual annotations. Segmented regions were subsequently analyzed using a second CNN classifier, which distinguished immunocompetent from immunodeficient tumors with 80% accuracy on an 80/20 train-test split. Tumors in RAG KO mice exhibited accelerated growth and absence of T-cell–driven infiltration, consistent with their underlying biology, while C57BL/6 tumors demonstrated features of adaptive immune activity. Together, these findings demonstrate the feasibility of ultrasound-based deep learning to stratify tumors by immune competence, offering a foundation for non-invasive immunological phenotyping and longitudinal treatment monitoring. |