Abstract
Cancer heterogeneity across spatial, temporal, genomic, and proteomic scales complicates diagnosis, prognostication, and therapeutic selection. Conventional biopsy‐based assays remain invasive, costly, and limited in capturing real-time tumour dynamics. Label-free, microfluidic approaches and recent advances in deep learning offer complementary routes for addressing these limitations; however, their combined diagnostic potential has not been comprehensively evaluated.This dissertation aims to develop and validate integrated microfluidic–deep learning systems for the label-free characterisation of solid tumours at cellular and sub-cellular resolutions. Specific objectives are: (i) to construct patient-derived tumour models on microfluidic chips that preserve three-dimensional architecture; (ii) to design computational pipelines for automated image analysis using convolutional neural networks, transformer-based architectures, and the Segment Anything Model; and (iii) to quantify metastatic potential through constriction-based deformability cytometry coupled with a dedicated analysis framework (ATMQcD).
Microfluidic devices were fabricated in polydimethylsiloxane by standard soft lithography. Patient tumour fragments from breast, lung, and bladder cancers were cultured in microwell arrays, imaged longitudinally, and processed using a hierarchical deep learning workflow that incorporated ResUNet++, Mask R-CNN, and vision transformers. For biophysical profiling, a high-throughput constriction-based deformability cytometer operating at ~25,000 cells min⁻¹ was integrated with ATMQcD, which performs automatic training-set generation, multi-object tracking, segmentation, and extraction of friction-related indices.
The findings demonstrate that coupling microfluidic tumour models with advanced deep learning analytics enables accurate, high-throughput, and cost-effective assessment of tumour heterogeneity and metastatic potential. The platforms developed herein offer a scalable foundation for precision oncology applications, particularly where rapid, label-free diagnostics are required.
| Date of Award | 9 Dec 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Bee Luan KHOO (Supervisor) |