(1) Background: Identifying early pancreas parenchymal changes remains a challenging radiologic diagnostic task. In this study, we hypothesized that applying artificial intelligence (AI) to contrast-enhanced ultrasound (CEUS) along with measurement of Heat Shock Protein (HSP)-70 levels could improve detection of early pancreatic necrosis in acute pancreatitis. (2) Methods: Acute pancreatitis
Acute pancreatitis (AP) is one of the most common gastrointestinal disorders, and cases have been steadily increasing over the past few decades (Yadav and Whitcomb,
The importance of early identification of AP severity may alter clinical decisions. In addition, speedy detection of pancreatic and/or peripancreatic necrosis, fluid collection, infection may change therapeutic protocols (Steinberg and Tenner,
Although computed tomography (CT) is considered a gold standard radiological modality for assessing local complications of AP, CEUS could be considered a safe alternative as it is a reliable, non-invasive imaging modality with no radiation exposure and high sensitivity and specificity. If the pancreatic region is clearly visible on ultrasound, CEUS could be used to assess the severity of acute pancreatitis It provides information on the pancreatic parenchyma’s vascularization, detecting areas of non-viable pancreatic tissue, i.e. necrosis. A significant correlation between CEUS and CT was found for pancreatitis CT severity index and the extent of necrosis for the severity of AP. As CT contrast medium is nephrotoxic, it can aggravate AP in animal models by impairing the pancreatic microcirculation (Foitzik
We have conducted a prospective observational study in the period between November 2018 and December 2020. All patients admitted to the Center of Abdominal Surgery, Emergency Room or Intensive Care Unit at Vilnius University Hospital “Santaros klinikos” (Lithuania) with a diagnosis of acute pancreatitis and onset of the disease within 72 hours were included in this study. Total number of 146
The diagnosis was established on the basis of acute abdominal pain, at least 3-fold elevated levels of serum amylase and typical radiological findings. The contrast-enhanced CT scan was performed on admission, on day 4 to 7 after onset of the disease, to demonstrate the presence of pancreatic necrosis. According to the clinical course and clinical severity scores (Modified Atlanta Classification for Acute Pancreatitis, APACHE II > 7, MODS > 2) patients were stratified into mild, moderate and severe acute pancreatitis groups. Clinical data related to the severity of disease, development of organ dysfunction and/or septic complications was prospectively collected in a standardized fashion. Patients with chronic pancreatitis or repetitive episodes of acute pancreatitis were excluded. Peripheral blood samples from patients were drawn on admission to the hospital. Afterwards, centrifugation serum HSP-70 levels were measured using the standard technique utilizing Human HSP-70 ELISA Kit (Bender MedSystems GmbH, Vienna, Austria).
HSP-70 concentration was quantified in the serum using a commercially available enzyme-linked immunoassay (Bender MedSystems GmbH, Vienna, Austria). In brief, human HSP 70 present in the sample binds to antibodies adsorbed to the microwells. Following incubation, unbound biological components are removed during a wash step and a biotin-conjugated anti-HSP70 antibody is added and binds to human HSP70 captured by the first antibody. Following incubation unbound biotin- conjugated anti-human HSP70 antibody is removed during a wash step. Streptavidin conjugated to horseradish peroxidase (HRP) is added and binds to the biotin conjugated anti- human HSP70 antibody. Furthermore, incubation unbound Streptavidin-HRP is removed during a wash step, and serum solution reactive with HRP is added to the wells. A coloured product is formed in proportion to the amount of human HSP-70 present in the sample. The reaction is terminated by addition of acid and absorbance is measured using a spectrophotometer at 450 nm. A standard curve is prepared from 7 human HSP70 standard dilutions and HSP70 sample concentration is determined. All specimens were tested in replicate wells.
CT scan shows the region of pancreas affected by AP, however, it is a more expensive technique and due to non-portability it is not suitable to be applied in operation room. Therefore, due to portability and lower application costs of CEUS technique, the related ultrasonic CEUS images were analysed further.
For CEUS imaging a non-invasive ultrasound imaging device (Hitachi Arietta 70, Hitachi Aloka Medical, Japan) and ultrasound contrast agent (SonoVue, Bracco, Milan, Italy) were used following the manufacturer’s instructions. To implement the AI-based classifiers for the automated classification of CEUS diagnostic images of clinical pathologies into the different classes, it is necessary to select, extract and calculate a set of quantitative parameters.
The acquired low mechanical index (0,07) contrast harmonic images (movie clips up to 60 seconds) were further post-processed by manually selecting the rough contour of two-dimensional (2D) spatial regions of interest (ROI) (Fig.
Calculations including image processing and extraction of the set of the quantitative parameters were performed using developed special algorithms implemented in the MATLAB 2020a software (The MathWorks Inc., Natick, MA USA). It is also essential to select and use only statistically significant parameters (
The workflow of post-processing of acquired CEUS images and AI-based classification for estimating the level of acute pancreatitis.
The principle of AI-based classifier for automatic estimation of the level of acute pancreatitis risk (none, low, moderate, or high) was presented as well (Fig.
The subjects were categorized into two groups (patient and healthy control groups) based on the presence of acute pancreatitis diagnosis on admission. The studied patient for an algorithm establishment was a 28-year-old male with severe acute pancreatitis who was admitted to the ER less than 24 hours after the symptoms started. The diagnosis was confirmed by clinical examination and by ultrasound examination. The patient’s condition became unstable, and he was transferred to the Intensive Care Unit (ICU), where a CT scan (Fig.
28 years old male with severe acute pancreatitis. Contrast-enhanced CT (portovenous phase, the same axial plane as on CEUS image) confirmed acute necrotizing pancreatitis (arrows) with acute necrotic collection (star).
The appropriate sets of quantitative parameters within the selected ROI of each CEUS images were acquired. Automated estimation of region boundaries of healthy parenchyma tissue within the pancreas ROI was performed by appropriate threshold analysis (e.g. Otsu, etc.). During the analysis, it was assumed that during CEUS procedure, the contrast material penetrates regions of healthy parenchyma tissue, with limited penetration in regions affected by necrosis of parenchyma tissue. Further, healthy and impaired areas (affected by necrosis) were automatically estimated and quantitatively evaluated as percentages compared of overall ROI (Figs.
Example of a CEUS image (15 s after injection of contrast agent) of a healthy pancreas (healthy volunteer – female) and estimated area of a healthy parenchyma (according to the presence of perfusion): a – contrast harmonic image of the pancreas region and manually selected ROI according to the solid red line, b – extracted informative ROI for further automatic detection of healthy parenchyma areas being marked with a solid green line, c – automatically detected area of a healthy parenchyma is binary marked (white) and covers 100.0% of the overall pancreas ROI.
A CEUS image of a pancreas at 22 s after injection of contrast agent, and estimated area of a healthy parenchyma (according to the presence of perfusion) from a male patient with acute necrotizing pancreatitis: a – contrast harmonic image of the pancreas region and manually selected ROI marked with a solid red line, b – extracted informative ROI for further automatic detection of healthy parenchyma areas marked with a solid green line, c – automatically detected area of a healthy parenchyma (white colour) covering 55.6% of overall pancreas ROI.
An example of a CEUS image of a healthy pancreas (healthy female volunteer) acquired at 15 s after injection of contrast agent and an estimated area of healthy parenchyma (according to the presence of perfusion) is presented in Fig.
Reconstructed normalized perfusion dynamic curves by approximations of log-normal distributions (Dietrich
Reconstructed normalized perfusion dynamic curves by approximations of log-normal distributions for healthy volunteer and patient with pancreas affected by necrosis.
In order to validate the reliability of non-invasive techniques (e.g. CEUS), the gold standard technique should be used as well. For example, measures in diagnosing acute pancreatitis are three times increased lipase and amylase levels. Recently, new and novel biomarkers have emerged. HSP-70 has been evaluated for use in the early diagnosis of acute pancreatitis. Bhagat
Estimated quantitative parameters from the reconstructed normalized perfusion dynamic curves.
Parameters | Parenchyma affected by necrosis (patient with acute pancreatitis) | Healthy pancreas of healthy volunteer |
Peek value, arbitrary units | 46.5 | 155.5 |
Time to peek after injection of contrast agent, s | 19 s | 24.2 s |
(10 s after appearance of contrast agent caused reflections) | (10.9 s after appearance of contrast agent caused reflections) | |
Area under the “wash in” curve | 7858 | 29595 |
Mean transit time MTT, s | 34.6 s | 40.1 s |
MTT (50% of peak), s | 23.2 s | 26.5 s |
“Wash in” rate | 9.3 | 28.7 |
“Wash out” rate | 2.1 | 5.7 |
Rise time, s | 8.4 s | 9.3 s |
Fall time, s | 24.6 s | 29.1 s |
Wash in perfusion index | 940.96 | 3175.1 |
In this paper we present two novel AI-based approaches combined with biochemical markers for early AP severity detection. First, automatic adaptive detection of ROI boundaries of the pancreas region within multiple contrast harmonic imaging (CHI) mode images. Second, the compensation of artifacts influences the stability of spatial pancreas tissue position within the detected ROI while comparing with HSP-70 levels. Additional artifacts are caused by physiological movements of patient body tissues due to breathing and pulsation of larger blood vessels and by additional transducer movements caused by the ultrasonography operator during examination and the image acquisition procedure.
Further study results are currently being analysed to process more images of healthy volunteers and patients with acute pancreatitis of various severity levels. Unfortunately, the worldwide outbreak of SARS-Cov-2 makes some limitations of clinical studies with real patients and healthy volunteers. Afterwards, the AI–based classification will be applied to detect the presence of early-stage AP severity. While CEUS remains a reliable non-invasive diagnostic test for detecting morphological changes of the pancreas, evaluating conditions of the vasculature, visualizing contours of pancreas, quantifying perfusion and necrotic areas, it could be used along with CT and magnetic resonance imaging (MRI) (Omoto
The importance of HSP-70 expression after CEUS have been described in liver diseases with the mean density 0.35 while compared in radiofrequency ablation group with 0.31 accordingly (Liu
In our study, the algorithm concept and workflow of post-processing of acquired CEUS images was used to estimate sets of quantitative parameters for healthy tissue of pancreas parenchyma and tissue affected by necrosis. Also, the principle of the AI–based classification for estimating the risk level of acute pancreatitis (none, low, moderate, high) was presented to overcome the limitations of ultrasound. Major disadvantages of ultrasound remain the limited visibility of the pancreas and peripancreatic region in a large proportion of patients with severe acute pancreatitis because of the presence of overlying bowel gas, particularly in the case of ileus. The body habitus may also limit the penetration of acoustic waves in obese patients. Additionally, abdominal ultrasound is less accurate in delineating extrapancreatic inflammatory spread within retroperitoneal spaces. Finally, ultrasound is operator-dependent and displayed on a limited number of images that are not easy to comprehend and convey to practicing clinicians. Applying AI and CEUS enhance sensitivity and specificity for the assessment of the early severity of acute pancreatitis.
Authors declare that the small number of patients and additional artifacts in the images are limitations to the study.
Further studies are now underway to determine whether pairing the CEUS, conventional radiology tests (i.e. CT, MRI) and AI techniques (image acquisition, further sophisticated post-processing, and classification) and biochemical markers (i.e.HSP-70) will result in a valuable new diagnostic and clinical decision support tool in early diagnosis of AP and the presence of pancreas tissue necrosis.
A patent application was filed describing the findings in this paper. The patent application number is LT2020 538.
We acknowledge the technical support and guidance of Ultrasound Research Institute, Kaunas Technology University, Kaunas, Lithuania.