Deep learning methods for reconstruction and analysis of diffuse optical tomography images of breast cancer lesions

The development of an accurate, efficient, portable, and affordable method for identifying breast cancer is critical for both early detection and improved prognosis. Medical imaging modalities play a critical role in cancer screening and treatment monitoring. Diffuse optical tomography (DOT) is a non-invasive imaging modality that can be used in a low-complexity probe design, resulting in an inexpensive portable imaging diagnostic device with low power consumption. In recent years, machine learning techniques have created transformative opportunities for medical image reconstruction and analysis, helping move toward data-driven algorithm designs wherein computational power is augmented with physics priors to push the accuracy and fairness of image driven diagnosis to new limits. In this thesis, we present multiple deep learning-based medical image reconstruction and analysis approaches for screening breast cancer lesions acquired by DOT. First, an end-to-end image reconstruction model from sensor-domain data is proposed, where physics-based simulation is leveraged to address the lack of available real-world data required for training. Next, we adopt a transfer learning strategy to align and translate the sensor domain distribution between in silico and real-world data and propose a novel loss to promote appearance similarity and penalize artifacts. Following up on this we propose a joint reconstruction and localization solution that simultaneously attends to the most important features while ensuring better lesion localization. Finally, we propose an orthogonal multi-frequency fusion solution for direct prediction of the end task from sensor signal data, increasing diagnosis accuracy at a reduced computational cost. Extending a portable device with such diagnosis ability promises to improve first-line treatment throughput. These contributions demonstrate the promising role of deep learning in DOT image reconstruction and diagnosis. Combined, our contributions open the path towards personalized medicine for non-invasive portable diagnosis and treatment monitoring of breast cancer in the very near future.