[2H5-GS-13] AI application: Image processing

[2H5-GS-13-05] Identifying the Corresponding CT Slices among the Different Scans via Deep Metric Learning

〇Barata Tripramudya Onggo¹, Joji Ota², Takuro Horikoshi³, Hajime Yokota⁴, Yasukuni Mori⁵, Hiroki Suyari⁵ (1. Chiba University Faculty of Engineering, 2. Chiba University Hospital Department of Radiology, 3. Chiba University Hospital Department of Radiology , 4. Chiba University Graduate School of Medicine, 5. Chiba University Graduate School of Engineering)

Keywords:Deep Metric Learning, Triplet Loss, CT Image

To observe the condition of a patient in long-term medical treatment, it is a common practice to compare computed tomography (CT) slices from current scans with existing CT slices from past scans of the same patient. A medical practitioner needs to identify the corresponding CT slices among the scans of different dates before being able to compare these CT slices, and the burden of identifying the right slice escalates as the number of follow-up examinations increase. Therefore, in this work, we aim to provide a methodology to identify the most similar CT slice from prior dates when a slice of interest is specified by applying Deep Metric Learning to quantify the similarity of CT slices captured on different scan dates. In the case studies, we showed that on a task to pinpoint the matching CT slices of different dates, a network trained with the proposed method can discover similar objects between slices and gives an output equivalent to visual assessment by an experienced radiologist.