Multimodal deep convolutional neural network based optical tomographic image reconstruction for enhanced contrast recovery in small breast lesions (Conference Presentation)

Near-infrared diffuse optical tomography (DOT) is a promising functional modality for breast cancer imaging; however, the clinical translation of DOT is hampered by technical limitations. Specifically, conventional finite element method (FEM)-based image reconstruction methods are time-consuming and ineffective in recovering full contrast, especially in small lesions typically seen in the diagnostic setting. To address this, in our prior work, a three-part deep convolutional neural network (DNN)-based model comprised of Fully connected layers, followed by a convolutional encoder-Decoder, and a U-Net (a.k.a. FDU-Net) was developed for fast, end-to-end 3D DOT image reconstruction. In this work, we further expanded FDU-Net to utilize high-resolution structural images as an additional input in a multi-channel U-Net. Through benchmark comparisons with conventional finite element-based methods (FEM) and previously published DNN models using 250 synthetic cases, the FDmU-Net demonstrated far superior accuracy in lesion localization and further enhancement in contrast recovery while maintaining high overall image quality in small 8-mm diameter inclusion cases.