Machine learning for inferencing of Hepatocellular Carcinoma (HCC) risk score from medical imaging

Early detection of Hepatocellular carcinoma (HCC) early is crucial as prognosis correlates with delays in diagnosis. HCC surveillance is important to identify tumors at an early, preclinical stage. However, there is significant variation in practice, often leading to inconsistent diagnosis and treatment decisions. To improve standardization and consensus regarding reporting HCC screening liver imaging, the LI‐RADS was launched in March 2011 by the ACR and adopted at Emory for Liver MRI reporting in 2018. During the transition phase from non-LI-RADS reporting to LI-RADS reporting, a difficult management question arises: how do we standardize report coding to support longitudinal patient follow-up when the patient’s screening imaging was reported before LI-RADS was implemented? Our project lies in the intersection of NLP, computer vision, and deep learning to build an effective AI tool that can infer the LI-RADS score directly from the HCC screening images without human interaction. The core novelty of the proposal is that no human-labeled data is required in any step of this study; for training the AI tool, we will use the LI-RADS MRI coded reports to train an NLP model to infer LI-RADS score on the historic reports. The proposed research will develop a powerful AI pipeline for calculating standardized scorings from radiological images which can be extended for other organs (e.g. lung, prostate) and will ultimately reduce practice variation between the radiologists and streamline radiology workflow.