Abstract CT274: Diagnosis based on signal : The first time break the routinely used circle of signal-to-image-to-diagnose

Abstract Purpose: The routinely used diagnostic scheme of cancers follows the process of signal-to-image-to-diagnosis. It is essential to reconstruct the visible images from the signal of medical device so that the human doctor can perform diagnosis. However, the huge amount of information inside the signal is not optimally mined, which causes the current unsatisfactory performance of image based diagnosis. In this study, for the first time, we developed an AI based diagnostic scheme for lung nodules directly from the signal (rawdata) to diagnosis, skipping the reconstruction step. Experimental Design: In this study, we focused on the discrimination of malignant from benign lung nodules using CT rawdata. We collected a simulation dataset (n=606, from LUNA public dataset; rawdata was generated from CT images) and a prospective dataset (n=268, from the First Hospital of Ji Lin University, China; Clinical trials: NCT04241614). In the prospective dataset, all patients underwent preoperative CT scan (both rawdata and CT images were available) and had pathologically confirmed result of the nodules. We used 3D-CNN (Convolutional Neural Networks) to generate a model from rawdata (Model_rawdata) and another model from CT image (Model_image). The performances of the two models were compared with area under curve (AUC). In addition, we explored the additional value of rawdata to image-based diagnostic scheme. Results: The simulation study showed that rawdata could distinguish malignant lung nodules (AUCs for training and validation cohorts were 0.88 and 0.80, respectively). In the prospective study using only the lung area, both Model_rawdata and Model_image had good classification performance (both AUCs were greater than 0.65). In the prospective study using only the nodule area, the Model_rawdata performed similar with Model_image (both AUCs were greater than 0.7). The stratification analysis showed that the Model_rawdata was not affected by CT scanning parameters. Furthermore, the addition of rawdata to image (Model_combination) could improve the diagnostic performance. Conclusions: Without reconstruction, the rawdata from CT system had comparative performance to image-based diagnosis. The rawdata could provide additional diagnostic value to images. This research breaks the routinely used circle of imaged based diagnosis, which may open up a new field for AI based diagnosis. Citation Format: Di Dong, Bingxi He, Boyu Kong, Lei Zhang, Lixia Tong, Feng Huang, Dong Han, Zhipei Huang, Huimao Zhang, Jie Tian. Diagnosis based on signal : The first time break the routinely used circle of signal-to-image-to-diagnose [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr CT274.