A Multimodal Longitudinal Generative Adversarial Network (GAN) to Discriminate High-risk Cysts for the Early Detection of Pancreatic Cancer

Pancreatic cancer (PC) is the deadliest common cancer with abysmal survival and a large unmet clinical need [1]. It is already the fourth leading cause of cancer death in Western countries; projected to be the second within a decade [2]. Early detection and diagnosis of PC may allow more treatment options to reduce this lethality to an extent by improving this dismal survival statistics. However, PC is hard to detect early due to: 1) low incidence of PC; 2) absence of well-validated high-risk groups; 3) vague and non-specific early symptoms; 4) absence of cheap and effective diagnostic tests; 5) complex PC biology and deep, inaccessible pancreatic anatomy hindering detection; 5) lack of validated diagnostic biomarkers; and 6) lack of sensitive imaging techniques for accurate detection of small pre-cancerous lesions. Pancreatic cysts (PANC-CYS) are considered precursor lesions to PC and are increasingly being recognised incidentally during scans performed for screening for abdominal aortic aneurysms or colorectal cancer [3]. However, most incidentally diagnosed PANC-CYS are benign and do not require further investigations, but currently there is no robust approach to distinguish the benign from pre-malignant and malignant cysts. As a result, expert consensus guidelines (e.g. Fukoka, European) are used for surveillance, resulting in not only significant healthcare burden (over-investigation and over-treatment) but more importantly, a major anxiety in these patients. Moreover, a number of other factors, most of which are routinely collected such as, but not limited to, demographic, lifestyle, medical conditions, pathology and blood/urine tests could potentially help in triaging patients better. Furthermore, emerging mining of imaging data, such as radiomics, may help in discriminating benign from pre-malignant lesions. PANC-CYS-GAN is an ambitious, ground-breaking, disruptive multi-modal artificial intelligence (AI) system aiming to revolutionize the early diagnosis of PC based on a holistic approach, powered by interdisciplinarity (computer vision, radiomics, statistics along with clinical guidance), using longitudinal data to mine meaningful hypothesis (e.g., a cyst from an ethnic minority male with smoking and drinking lifestyle and family history of diabetes has 70% chance of malignant transformation in next 2 years). We hope to identify the possible predictors of PC in a non-invasive manner for efficient diagnosis, resulting in improved early detection. Successful results from this pilot project would potentially revolutionize the early detection and diagnosis of PC, resulting in significant health, social and economic benefits to the UK, and beyond.