2

What we learned from a decade of research on imaging of CP?

Temel Tirkes, MD, FACR, FSAR

Professor of Radiology & Imaging Sciences, Medicine and Urology

Chronic Pancreatitis - What Have We Learned From a Decade of Prospective MRI Research and AI?

​

Temel Tirkes, MD, FSAR, FACR 

Professor of Radiology & Imaging Sciences, Medicine and Urology

Ulas Bagci, PhD

Director of Machine & Hybrid Intelligence Lab

Disclosures

NIH support

R01-DK116963 - Magnetic resonance Imaging as a Non-Invasive Method for Assessment of Pancreatic fibrosis (MINIMAP) study

U01-DK127382 - Type 1 Diabetes in Acute Pancreatitis Consortium

U01-DK108323 - Consortium for the Study of Chronic Pancreatitis, Diabetes, and Pancreatic Cancer

R01-CA260955 - Predicting Pancreatic Ductal Adenocarcinoma (PDAC) Using Artificial Intelligence Analysis of Pre-diagnostic Computed Tomography Images

R01-DK132631 - A Pilot Clinical Trial of Paricalcitol for Chronic Pancreatitis

​

Financial conflict: None

​

atirkes@iu.edu

3

atirkes@iu.edu

4

Publications on CP past 10 years

atirkes@iu.edu

5

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

Correlation of MRI Parenchymal Features with Histopathology

MRI T1 signal vs MRCP Cambridge Score

Histopathology MRI T1 Signal Ductal Imaging

Abdom Radiol (NY). 2022 Jul;47(7):2371-2380. PMID: 35486166; Tirkes et. al.

MRI T1 Score and Fibrosis

r = 0.54

MRCP Cambridge score and Fibrosis

r = 0.26

MRI T1 signal vs MRCP Cambridge Score

Abdom Radiol (NY). 2022 Jul;47(7):2371-2380. PMID: 35486166; Tirkes et. al.

Correlation of MRI with fibrosis

r = 0.44

Am J Gastroenterol 2015; 110:1598–1606; doi: 10.1038/ajg.2015.297; September 2015

57 Adult patients underwent TPIAT for CP

atirkes@iu.edu

10

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

MRI and MRCP in suspected CP

T1 Signal in suspected CP

T1 SIR

RAP patients suspected of CP, but with normal MRCP

Lower T1 signal reflects low HCO₃

Normal HCO₃

Abnormal HCO₃

CP before any ductal changes

atirkes@iu.edu

14

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

PROCEED Study

MRI, MRCP and CT on 2,000 RAP/CP patients

T1 Score (SIR of pancreas to spleen)

No CP

CP

Fat is suppressed. No contrast given. Performed in almost all MR examinations

T1 Score (SIR) �is an imaging biomarker of CP

atirkes@iu.edu

17

• 820 MRIs in well-phenotyped control, RAP and definite CP
• Included all MR vendors
• Included 3.0T and 1.5T

​

​

T1 Score decreases with increasing CP severity

​

Abdom Radiol (NY). 2022 Oct;47(10):3507-3519. PMID: 35857066. Tirkes, ae. al.

19

T1 Score

T1 Score correlates with severity of CP

T1 score of 1.2 is the threshold

atirkes@iu.edu

20

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

Longitudinal Follow-up of CP

• Patients with definitive CP had imaging follow-up of 4 years.
• MR imaging parameters:
• Pancreatic volume
• DWI
• Fat signal fraction
• MPD diameter

MRI vs MRCP for Follow-up of CP

Morphological progression of CP seems to be primarily parenchymal related

{

Pancreas volume decreased

Fat increased

MRI diffusion decreased

Ducts did not change

{

atirkes@iu.edu

23

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

Discordance of ERCP and MRCP interpretation

325 CP patients had MRCP and ERCP within 90 days

MRCP and ERCP based agreement of Cambridge score was only 43%

​

Conclusion

There is no satisfactory concordance between ERCP and MRCP-based Cambridge scores

atirkes@iu.edu

25

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

Inter-observer agreement of Cambridge �by CT and MRCP

MR kappa = 0.68

CT kappa = 0.56

Composite CT/MR Kappa = 0.62

MRI parenchymal features IOV

atirkes@iu.edu

27

Cambridge Classification - Weighted Kappa: 0.54

atirkes@iu.edu

28

MR Imaging Biomarkers for CP

MSCP 0 1 2 3 4 5 6

​

T1 Score 1.34 1.27 1.21 1.16 1.19 1.12 1.05

T1 ms (1.5T) 400 500 600 700 800

ECV (%) 10 20 30 40 50

Volume (ml) 80 70 60 50 40

AVR (art/ven) 1.4 1.3 1.2 1.1 1.0

Tail dm (cm) 2.5 2.4 2.3 2.2 2.1

​

CP-MRI Score 0 1 2 3 4 5

​

atirkes@iu.edu

29

​

​

​

​

What have we learned

from a decade of research on

MR imaging of CP?

atirkes@iu.edu

30

Normal

CP

Pancreatic steatosis and CP

MRI Fat Quantification

Fat only image Water only image

Pancreas

PDFF is better than dual echo DIXON

Pancreatic Fat and Adipose Tissue Quantification

Manual AI

Visceral and subcutaneous adipose tissues can be separated and measured either manually or AI tools

Pancreatic steatosis, Obesity, and CP

33

?

atirkes@iu.edu

35

Presenter: Prof. Ulas Bagci

Northwestern University, Machine and Hybrid Intelligence Lab, February, 2025

AI in Pancreas Imaging & Analysis

Disclosures for Dr. Bagci

NIH support

​

R01-CA246704- Cyst-X: Interpretable Deep Learning based Risk Stratification of Pancreatic Cystic Tumors

​

U01-DK127384-02S1- Data Coordinating Center for the Type 1 Diabetes in Acute Pancreatitis Consortium-Imaging Morphology of Pancreas in Diabetic Patients following Patients with Acute Pancreatitis (IMMINENT) Study

​

​

Financial conflict: None.

​

atirkes@iu.edu

36

Machine and Hybrid Intelligence Lab

• Founded in 2021 January
• School of Medicine, Dept of Radiology
• Courtesy Biomedical Engineering, 
• Courtesy Electrical and Electronics Eng, 
• PI @ Lurie Cancer Center
• Member – NIH AIR (AI in Resource)
• Part of I.AIM (AI Institute at Northwestern)
• 3 Large-scale servers (>30GPU)
• ~30 members

​

www.bagcilab.com

37

38

AI is needed for Pancreas Imaging Research for three primary tasks

1. Standardization / Harmonization of Images
• Heterogeneity of MRIs, scanner and acquisition differences, patient profiles, etc.
2. Segmentation of Pancreas from radiology images (CT, MRI,...)
• Compared to other organs, very challenging problem due to complex shape, sparse position, changing orientation, and poor signal in certain boundary locations
3. Prediction algorithms
• Diagnosis (a patient has this condition or no?)
• Patient outcome prediction (a patient will have this condition or no?, will they respond to therapy? and similar questions)

39

General Overview – How to Use AI in MRI Research of Pancreas

40

General Overview – How to Use AI in MRI Research of Pancreas

41

General Overview – How to Use AI in MRI Research of Pancreas

42

General Overview – How to Use AI in MRI Research of Pancreas

43

CP --> severity

​

or

​

No? --> future risk?

1. Harmonization / Preprocessing

atirkes@iu.edu

44

2. Pancreas & Its Part Segmentation

• Historically very difficult problem, most existing algorithms focus on CT imaging
• Pancreas volumetry and its parts (head, body, tail) are critical to measure
• Segmented pancreas can be used for auto-diagnosis and future-prediction with AI

45

46

PanSegNet

Summary of Contributions for PanSegNet

• First ever multi-center, muti-site MRI pancreas segmentation algorithm
• State of the art results are obtained!
• 3D (volumetric) Algorithm
• Code, Models, Data, and Labels are made public
• Largest Cohort reported so far (>700 MRI with pubs, currently >1400 with updated version)
• CP, Cyst, Diabetes, pediatric pancreatitis patients were all tested (>2000), it works also for PanCan patients CT (>2000)

​

​

47

Inter- and Intra-observer Analysis of GT Labels

48

CT Results (>88% dice accuracy)

49

MRI T1 Results

50

MRI T2 Results

51

52

Out-of-Phase MRI Segmentation (>87% dice score)

53

Pancreas Part Segmentation – Head/Body/Tail��>87% dice score

54

Pediatric patients – pancreatitis (acute/chronic)

55

Tumor / Cyst Segmentation ?

atirkes@iu.edu

56

MICCAI AIPAD 2024

3-AI and Radiomics for Diagnosis and Prognosis

atirkes@iu.edu

57

Radiomics

Radiomics

Radiomics

Classification of Chronic Pancreatitis (with Radiomics)

Frøkjær et. Al., Abdominal Radiology (2020) 45:1497–1506

n=77 CP patients and 22 healthy controls (using DWI)

​

CP: 98% accuracy

​

Diabetes: 83% accuracy

​

EPI: 82% accuracy

​

​

Classification of Pancreatic Diseases

After segmentation, region of interest data is fed into the classifier (deep learning) to identify the pathologies

https://www.mdpi.com/2072-6694/14/2/376

63

2/17/2025

IEEE EMBC 24

https://arxiv.org/pdf/2404.17064

​

Classification of Pancreatic Cancer

Deep learning and radiomics show strong potential for improving pancreatic cancer diagnosis using CT and MRI.

​

2024

atirkes@iu.edu

65

Future Steps

• Multimodal Deep Learning

Combination of imaging, clinical, pathological, and other biomarkers

​

• Explainable / Trustworthy AI

AI should have reasonings/explanations for their predictions

​

• Multicenter trials

AI has generalization problems, one center data-based model may not work in other centers

Challenges

• Diagnosis / Prognosis requires explainability

atirkes@iu.edu

67

Challenges

• Data sharing & Multi-Center Studies --> ethical / privacy issues

atirkes@iu.edu

68

Challenges

• Data sharing & Multi-Center Studies --> ethical / privacy issues

atirkes@iu.edu

69

Challenges

• Data sharing & Multi-Center Studies --> ethical / privacy issues
• IEEE ISBI 2025 (H. Pan et al)

atirkes@iu.edu

70

Dice Jaccard Precision Recall HD95 ASSD

ACC AUC ACC AUC

T1 T2

Thank you for listening, Questions?

​

atirkes@iu.edu

71