Forward or backward? comparison of direct inversion and forward optimization techniques for DL-DOT

Diffuse Optical Tomography (DOT) is an imaging method that utilizes near-infrared light to image the human body. The image is acquired by inverting a propagation model based on measuring the scattered fields. One of the most promising methods to solve DOT is deep learning, due to its high computational power and fast inference. However, the ill-posed nature of the problem and the non-uniqueness of its solution can pose challenges in learning the inverse transformation. To address these challenges, two main approaches are used: learning the forward mapping and then applying classical inversion techniques, or directly learning the inverse mapping. In this study, both approaches were applied on a realistic breast model to evaluate the impact of dataset size on the optimal technique for deep learning-based Diffuse Optical Tomography. The results showed that for the specific scenario the direct inversion approach was superior for large datasets (over 2K), improving the RMSE and CNR by 44% and 97% respectively. Conversely, the forward optimization approach was better for smaller datasets (150), improving both the RMSE and CNR by 10% and 67% respectively.