A tomography method, system, and apparatus based on time-domain spectroscopy are provided. A light emitter is controlled to emit a pulse beam to scan a cross-section of an object to be measured while using a light receiver to detect the pulse beam passing through the object to be measured, so as to obtain time-domain pulse signals at locations of a scan path. A scan angle is repeatedly changed to perform the scanning and detecting steps, so as to collect the time-domain pulse signals of multiple angles of the cross-section as a time information set. Features are retrieved from the time-domain pulse signals using kernels of a trained machine learning model, which is trained with time information sets and corresponding ground truth images of cross-sections to learn the kernels for retrieving the features. The retrieved features are converted into a spatial domain to reconstruct a cross-sectional image of the object. |