Photo Magnetic Imaging is an imaging technique first proposed and created at the Center for Functional Onco Imaging at the University of California, Irvine. This imaging technique was first developed to early detect and identify cancerous regions in the biotissue using a combination of near-infrared laser-based optical imaging and Magnetic Resonance Thermometry (MRT) techniques. In PMI, the imaged tissue is illuminated and slightly warmed up with a near-infrared laser. The laser-induced temperature increase through absorption is measured using MRT. PMI absorption images are then obtained using a multiphysics solver combining light and heat propagation. This model is used to describe the spatiotemporal distribution of laser-induced temperature increase. Then, a dedicated PMI reconstruction algorithm is used to recover high-resolution optical absorption maps from temperature measurements. Being able to perform measurements at any point within the medium, PMI overcomes the limitations of conventional diffuse optical imaging. Higher absorption regions warm up more than the tissue background and thus can be directly seen on the MRT maps. These regions are observed as highly diffused bright blobs due to temperature diffusion. In fact, the boundaries of these regions are generally smaller than these diffused blobs. Therefore, overcoming the diffusion effect and accurately detecting the real boundaries of these regions provides invaluable information about their location and size. The focus of this research work is to provide a fast computational technique to employ novel machine learning techniques to perform the challenging aforementioned task. Here, we propose a Region-Based Convolutional Neural Network methodology in which the MRT temperature maps are first decomposed into square overlapping regions that cover the whole imaged medium. Then, Convolutional Neural Networks are used to classify these regions into 1) Positive, if they contain one of these diffused blobs, or 2) Negative, if not. Finally, a multi-regression model is used to identify the boundaries of these regions using an intersection over union of all the Positive square regions our method is tested and validated on a variety of cases with a mean accuracy of 95.35%.