The diagnosis of pulsatile tinnitus (PT) typically relies on medical imaging tests. However, due to the wide range of possible causes, there is still a lack of universally accepted diagnostic criteria with a clearly defined mechanism. This study aims to propose a neural network model for high-accuracy PT identification based on CT images of PT patients and non-PT individuals, as well as automatically label the high pathogenic regions to assist in diagnosis. Transfer learning based on the ResNet-v1-50 model was employed to identify PT using horizontal cross-sections of the middle temporal bone in the bone window. The high-weight regions for identification were labeled using the grad-CAM method. These regions, along with related anatomical structures, were statistically analyzed across three databases: large sections (entire cranium), medium sections (bilateral temporal bones), and small sections (right-side temporal bone), allowing for the gradual refinement of the area of interest and increased resolution of high-weight regions in classification. The best identification, which achieved 100% accuracy in the test set, came from the medium area that included both temporal bones. The high-weight regions identified in PT were concentrated in either bilateral or unilateral temporal bones, primarily involving the temporal bone air cells, tympanic antrum, sigmoid sinus cortical plate, and superior tympanum. The occurrence of PT is closely associated with temporal bone and nearby bone structures. PT patients have a probability of structural abnormalities in either bilateral or contralateral temporal bones, which are different from those without tinnitus. Specifically, bone structures including temporal bone air cells, tympanic antrum, sigmoid sinus cortical plate and tympanic cavity have a high probability of containing the primary pathogenic factors for PT. These imaging-based conclusions align with previous biomechanical findings, further corroborating the understanding of PT etiology.