Intra-pancreatic fat deposition (IPFD) refers to lipid accumulation within the pancreas, which manifests as adipocytes or lipid droplets within acinar and endocrine cells. IPFD is a modifiable condition that can be improved through dietary and exercise interventions. High IPFD is typically identified by increased pancreatic echogenicity on ultrasonography and can be quantified using computed tomography or magnetic resonance imaging. Approximately 20% of the general population exhibits IPFD levels exceeding 10% of the pancreas, making it one of the most common pancreatic conditions. IPFD shows a modest correlation with subcutaneous, visceral, and hepatic fat, and can occur in non-obese individuals. Genome-wide association studies have identified genetic polymorphisms linked to IPFD, while environmental factors such as aging and obesity also play important roles. Recent studies suggest that IPFD is a risk factor for pancreatic cancer, pancreatitis, and diabetes. However, high-quality clinical epidemiological research integrating pathology, medical imaging, clinical data, and genetic information is needed to clarify these causal relationships.

Pancreatic cancer is one of the malignancies with the poorest prognosis among all cancer types. However, it is believed that a cure is achievable if detected at the stage of carcinoma in situ or as a microlesion less than 10 mm in size. Currently, only 3% of pancreatic cancers are detected at 20 mm or less (Stage I or earlier). Therefore, the establishment of an early screening method for pancreatic cancer is an urgent issue. In collaboration with Fujifilm Corporation, we have developed an AI-based image diagnostic support technology that automatically recognizes not only direct findings, such as tumor identification, but also indirect findings suggestive of pancreatic cancer, including pancreatic duct dilation, stenosis, and pancreatic atrophy. This AI-based diagnostic support technology is expected to contribute to improving prognosis by increasing the detection rate of pancreatic cancer during screenings.

Pancreatic diseases are challenging to diagnose definitively with only CT or MRI, and procedures such as endoscopic ultrasonography (EUS) and EUS-guided fine needle aspiration (EUS-FNA) are necessary. However, these procedures are highly operator-dependent, making it challenging to standardize their quality. Artificial intelligence (AI) has started to be utilized in medical image diagnosis and is becoming integrated into clinical practice. While no AI models specifically for pancreatic diseases are approved in Japan, there have been several research reports on AI for detecting and differentiating pancreatic tumors and cysts. Since pancreatic disease prevalence is lower than gastrointestinal diseases, the cost of collecting the data necessary for AI development is higher, which has delayed progress in this field. It is necessary to establish efficient data collection methods and learning strategies to promote AI development for diagnosing not only pancreatic diseases but also other low-prevalence conditions.

Most histological types of malignant pancreatic tumors are adenocarcinomas. Therefore, understanding the characteristics of these adenocarcinomas is crucial. However, tissue specimens obtained using endoscopic ultrasound-guided fine-needle biopsy (EUS-FNAB) tend to be small and prone to mechanical damage, thereby presenting challenges in terms of diagnostic accuracy. Furthermore, since EUS-FNAB is performed through a transgastric or transduodenal approach, contamination complicates the histological diagnosis. In recent years, the development of artificial intelligence (AI), which is being implemented in clinical settings, has advanced the field of pathology. AI applications are being explored for pancreatic pathological diagnosis. In this article, we summarize the diagnostic system developed by our research group and discuss the recent advancements in AI applications for pathological diagnosis.

Pancreatic cancer has an extremely poor prognosis, and the completion of preoperative therapy and successful resection significantly impact patient survival. In this study, we aimed to construct a machine learning model to predict the likelihood of achieving pancreatic cancer resection by utilizing clinical information available prior to the initiation of neoadjuvant chemoradiotherapy in patients with borderline resectable (BR) and unresectable locally advanced (UR-LA) pancreatic cancer. Using a random forest algorithm on data from 360 patients, we achieved a predictive performance with an area under the curve of 0.75. Analysis of feature importance and explainable AI (XAI) revealed that age, platelet count, maximum tumor diameter, white blood cell count, nutritional immune indicators, and superior mesenteric artery invasion, which are factors not fully captured by traditional statistical methods, were useful for prediction. These findings suggest the utility of machine learning in predicting the achievement of pancreatic cancer resection by capturing complex interactions and nonlinear relationships among features. Future efforts to expand data and integrate multifaceted features are expected to enhance the performance of predictive models for optimizing individualized therapy.

Background: We have investigated the usefulness of surgical simulation using an artificial intelligence (AI) engine designed with deep learning algorithms for pancreatic cancer patients.

Methods: 1. In 100 pancreatic cancer patients, tumor size, position, and stagewise correlations with the pancreatic parenchymal Dice coefficient (DC) were analyzed. The relationship between the pancreatic duct diameter and the DC, and between the manually and AI-measured diameters of the pancreatic duct were analyzed. 2. Among sixty pancreatic cancer patients (head 36, body 14, tail 10) with a DC of 0.8 or higher, the correlation between resected pancreatic volume (RPV), pancreatic resection surface area (PRSA), thickness of pancreatic parenchyma (TPP), pancreatic duct diameter (PDD) at the resection location and tumor position were investigated.

Results: 1. A positive correlation (r=0.61, P<0.001) was observed between the manually and AI-measured diameters of the pancreatic duct. 2. The RPV (%) by position (head: body: tail) was 79.9: 26.5: 30.2, respectively, and the resection rate for head lesions was significantly higher (P<0.001). The PDD (mm) by position (head: body: tail) was 8.3: 4.6: 3.9, respectively, and the PDD in head lesions was significantly higher (P<0.001).

Conclusions: AI engine was found to be useful as a surgical simulation tool for pancreatic cancer patients.

A 52-year-old man, who had been monitored for the last two years due to a pancreatic cystic lesion in the tail of his pancreas, was referred to our hospital after a recent MRI revealed a new cystic lesion measuring 45 mm in the same location. A contrast-enhanced CT scan showed low density inside the cystic lesion. Endoscopic ultrasound identified a papillary mass inside the cyst at the tail of the pancreas and an irregular hypoechoic mass lesion on the pancreatic head side. Furthermore, endoscopic retrograde pancreatography showed disruption of the main pancreatic duct in the tail of the pancreas. As the possibility of pancreatic cancer could not be ruled out, a laparoscopic distal pancreatectomy was performed. The tumor lesion was brownish-yellow in color, with clear margins and contained muddy deposits. Histopathological analysis identified cholesterol crystals surrounded by foreign body giant cells, and the tumor was ultimately diagnosed as a pseudocyst accompanied by cholesterol granuloma.

A 69-year-old man was admitted to our hospital with a diagnosis of acute pancreatitis. Computed tomography (CT) revealed a hypodense mass in the pancreatic tail, and magnetic resonance imaging (MRI) showed high- and low-signal intensity on T1- and T2-weighted images, respectively. Endoscopic ultrasound-guided tissue acquisition (EUS-TA) and serial pancreatic-juice aspiration cytologic examination did not reveal any evidence of malignancy. The patient refused surgical treatment and underwent follow-up with imaging. MRI performed 2 months later revealed shrinkage of the pancreatic tail mass without any significant change in the imaging findings thereafter. Imaging performed 30 months later showed local enlargement of the mass; therefore, we repeated a third EUS-TA, and the patient was diagnosed with a pancreatic neuroendocrine tumor (NET). We performed laparoscopic distal pancreatectomy, which confirmed a final diagnosis of NET G1. The delayed diagnosis (30 months) was attributable to the atypical imaging findings and course of the NET, and the lack of conclusive histopathological diagnosis at the time of initial examination.