Date of Award
10-22-2024
Degree Name
Doctor of Philosophy (PhD)
Degree Type
Dissertation
Abstract
This dissertation investigates the application of advanced deep learning techniques to enhance biomedical imaging technologies, specifically focusing on Diffuse Optical Tomography (DOT) and Photoacoustic Tomography (PAT) for the diagnosis and treatment of ovarian and breast cancers. By integrating convolutional neural networks (CNNs) and other machine learning architectures, the research addresses critical challenges in image reconstruction, 3D rendering, and classification tasks within medical imaging domains. For DOT, a novel machine learning model with physical constraints (ML-PC) was developed, significantly improving the quality and accuracy of image reconstructions from diffuse optical signals. This model effectively managed the inherently noisy and indirect measurements typical in DOT, providing clearer, higher-resolution images that enhance the detection and monitoring of medical conditions such as breast cancer. In the study of PAT, an ultrasound-enhanced Unet model was introduced, leveraging ultrasound features to enhance the reconstruction of optical absorption distributions in ovarian lesions. This model achieved high accuracy and diagnostic performance, effectively differentiating between malignant and benign ovarian lesions. The use of quantitative PAT, combined with the ultrasound-enhanced Unet model, provided a powerful tool for non-invasive imaging, offering significant improvements in detecting and characterizing ovarian cancer compared to traditional imaging methods. The PA-NeRF model addressed the limited-view problem in PAT, which is crucial for providing high-quality 3D reconstructions necessary for ovarian cancer diagnosis. By utilizing neural radiance fields (NeRF) for 3D PAT reconstruction from limited B-scan data, the model produced detailed and accurate 3D images from fewer scans. This advancement is particularly important for clinical applications where limited data acquisition is common, enabling better visualization and assessment of ovarian lesions and improving diagnostic capabilities. The USDOT-Transformer model was developed to predict pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in breast cancer, integrating multimodal imaging data for a robust assessment of treatment response. This model combined ultrasound and diffuse optical tomography data with transformer-based deep learning, providing valuable insights into the effectiveness of NAC and demonstrating significant potential for personalized treatment planning. These advancements highlight the potential of combining machine learning with optical imaging modalities to improve diagnostic accuracy and treatment planning in cancer care. The integration of deep learning techniques has shown to be particularly beneficial in overcoming challenges associated with traditional imaging methods, providing more reliable and detailed insights into tissue properties and treatment responses. This dissertation illustrates the transformative impact of deep learning on biomedical imaging, setting the stage for future innovations that could further revolutionize this critical area of healthcare technology.
Language
English (en)
Chair
Quing Zhu
Committee Members
Abhinav K. Jha; Adam Bauer; Neal Patwari; Song Hu