Deep Learning Exhibits ❯ RSNA 2019

Pancreatic Cancer Imaging: A New Look at an Old Problem

 

 

Pancreatic Cancer Imaging: A New Look at an Old Problem

Linda C. Chu, Seyoun Park, Satomi Kawamoto, Alejandra Blanco, Shahab Shayesteh, Saeed Ghandili, Daniel F. Fouladi, Alan L. Yuille, Ralph H. Hruban, Kenneth W. Kinzler, Bert Vogelstein, and Elliot K. Fishman

The Russel H. Morgan Department of Radiology and Radiological Science, Department of Computer Science, Department of Pathology, and the Department of Cancer Research, Johns Hopkins University, Baltimore

 

Disclosure:

Research support from The Lustgarten Foundation:
  • Linda C. Chu
  • Seyoun Park
  • Satomi Kawamoto
  • Alejandra Blanco
  • Shahab Shayesteh
  • Saeed Ghandili
  • Daniel F. Fouladi
  • Alan L. Yuille
  • Kenneth Kinzler
  • Bert Vogelstein
  • Elliot K. Fishman

 

Overview

Research support from The Lustgarten Foundation:
  • Overview of current state-of-the art pancreatic cancer imaging
  • Current areas of need and potential solutions
  • Potential applications of emerging techniques in pancreatic cancer imaging:
    • Radiomics
    • Deep learning
    • Cinematic rendering

 

Introduction

  • Pancreatic ductal adenocarcinoma (PDAC) is the 3rd most common cause of cancer death in the US, with dismal five-year survival of 8.2%
  • CT is the first-line imaging modality for the initial diagnosis and staging of PDAC
  • Accuracy of PDAC detection and determination of resectability critically depend on imaging technique and experience of the radiologists
  • Society of Abdominal Radiology and the American Pancreatic Association have consensus recommended CT protocol and reporting template
https://seer.cancer.gov. Chu LC et al. Cancer J. 2017;23(6):333.342

 

Current State-of-the-Art

Current State-of-the-Art

Al-Hawary MM et al. Radiology. 2014;270:248-60.

 

Reporting Template

Reporting Template

Al-Hawary MM et al. Radiology. 2014;270:248-60. Zaky AM et al. RadioGraphics. 2017;37:93-112.

 

Current Frontiers for Improvement

  • We have achieved high accuracy for detection and assessment of resectability with state-of-the-art CT imaging
    • CT accuracy for detection: 76-96%
    • CT accuracy for resectability: 73-89%
  • However, there are several frontiers we can improve upon:
    1. Patient prognostication to develop personalized risk assessment
    2. Computer aided diagnosis to reduce misdiagnosis
    3. Advanced visualization techniques to facilitate pretreatment planning
Chu LC et al. Cancer J. 2017;23(6):333.342.

 

Current Frontiers for Improvement

Emerging imaging and analytic techniques have the potential to improve pancreatic cancer imaging

Current Frontiers for Improvement

 

Patient Prognostication

  • Surgical resection is the only curative treatment for patients with PDAC, but it is a major surgery with its associated morbidities
  • Some of the currently known prognostic features are based on pathologic features that are available AFTER surgery (TNM stage, resection margin)
  • Improving patient selection using imaging biomarkers available BEFORE surgery is critical in identifying patients most likely to benefit from surgical resection
Brennan MF et al. Ann Surg. 2004;240:293-298. Bilici A. World J Gastroenterol. 2014;20(31):10802-10812.

 

Radiomics

  • Radiomics converts imaging data into high dimensional mineable feature space
  • Potential to provide imaging biomarkers to quantify tumor heterogeneity and predict biologic behavior
  • Radiomics features can be classified into first-order, shape, texture, and higher-order statistical outputs
Kumar V et al. Magn Reson Imaging. 2012;30(9):4-48. Aerts HJ et al. Nat Commun. 2014;5:4006.
Gillies RJ et al. Radiology. 2016;278(2):563-577.

 

Radiomics

Radiomics

Kumar V et al. Magn Reson Imaging. 2012;30(9):4-48. Aerts HJ et al. Nat Commun. 2014;5:4006.
Gillies RJ et al. Radiology. 2016;278(2):563-577.

 

Survival Predication – Our Experience

Retrospective study of 125 patients with surgically resected PDAC and preoperative dual-phase CT between 2010 to 2014:
  • Patient age
  • Gender
  • Baseline CA 19-9
  • Tumor size
  • 489 radiomics features
All the preceding factors predict patient survival.

 

Survival Predication – Our Experience

  • Patients were stratified based on low-risk (survival >900 days) and high-risk groups (survival <350 days)
  • Feature reduction by minimum-redundancy maximum-relevance selection and random forest method
  • 10 most relevant radiomics features were selected
  • Survival analysis was performed based on clinical parameters with or without the 10 radiomics features

 

Survival Predication – Our Experience

10 radiomics features showed 82% accuracy in classification of high-risk vs. low-risk survival

Radiomics features from preop CT may be useful in risk stratification prior to planned resection of PDAC

Survival Prediction - Our Experience

Image: Predicted Survival Curve Based on Clinical Parameters + 10 Radiomics Features

 

Survival Prediction – From the Literature

Survival Prediction - Our Experience

Radiomics features were additive to clinical features in predicting patient survival

Cassinotto C et al. Eur J Radiol. 2017;90:152-158. Eilaghi A et al. BMC Medical Imaging. 2017;17(1):38. Chakraborty J et al. PLoS One. 2017;12(12):e0188022. Yun G et al. Scientific Reports. 2018;8:7226. Attiyeh MA et al. Ann Surg Oncol. 2018;25:1034-1042.

 

Pancreatic Cyst Risk Stratification – Our Experience

  • Pancreatic cysts are identified incidentally in approximately 2% of abdominal CTs
  • Malignant potential of pancreatic cystic lesions vary based on underlying pathologic diagnosis
  • These cystic lesions can have overlapping imaging features and can be difficult to differentiate clinically
Laffan TA et al. AJR. 2008;191:802-807. Zanini N et al. Pancreatology. 2015;15:417-422.

 

Pancreatic Cyst Risk Stratification – Our Experience

Utility of radiomics in classification of pancreatic cystic neoplasms:
  • Retrospective study of 214 pancreatic cysts:
    • 64 IPMNs
    • 33 Mucinous cystic neoplasms
    • 60 Serous cystadenomas
    • 24 Solid pseudopapillary neoplasms
    • 33 Cystic pancreatic neuroendocrine tumors
  • Compared diagnostic performance of radiomics and random-forest classifier vs. a radiologist

 

Pancreatic Cyst Risk Stratification – Our Experience

  • Academic radiologist with >20 years experience:
    • Accuracy = 77.1%
  • Radiomics features with random-forest:
    • Accuracy = 82%
Radiomics features can achieve improved accuracy in classification of pancreatic cystic neoplasms compared to an experienced radiologist

Pancreatic Cyst Risk Stratification - Our Experience

 

Risk Stratification of IPMNs – From the Literature

Risk Stratification of IPMNs – From the Literature

Hanania AN et al. Oncotarget. 2016;7(52):85776-85784. Permuth JB et al. Oncotarget. 2016;7(52):85785-85797. Attiyeh MA et al. HPB (Oxford). 2019;21(2):212-218.. Chakraborty J et al. Med Phys. 2018;45(11):5019-5029.

 

PDAC: Opportunity for Earlier Detection

  • Although the reported accuracy of detection of PDAC from CT scans is >95%, the diagnosis remains challenging and many cases are missed
  • 18.7% of patients referred to a pancreatic cancer center of excellence had a change in the diagnosis and/or clinical stage
  • In a retrospective study, early findings of PDAC can be detected up to 34 months prior to clinical diagnosis of PDAC
Chu LC et al. Cancer J. 2017;23(6):333.342. Pawlik TM et al. Ann Surg Oncol. 2008. 15(8):2081-8. Gonoi W et al. Eur Radiol. 2017. 27(12):4941-4950.

 

PDAC: Opportunity for Earlier Detection

  • Opportunity to use computer aided diagnosis as a “second-reader” to reduce the number of misses
  • This can detect disease at an earlier stage, which will significantly improve patient outcomes
  • Traditional computer assisted diagnosis (CAD) systems depend on human-designed features
  • Newer CAD systems based on deep learning and artificial intelligence have achieved superior results compared to traditional CAD
Geras KJ et al. Radiology. 2019 [Epub ahead of print]

 

Deep Learning

Deep learning uses training data and multiple layers of equations to develop a mathematical model that fits the data

Deep Learning

Erickson BJ et al. RadioGraphics 2017;37:505-15. Kohli M et al. AJR 2017;208:754-60

 

Deep Learning

  • Currently, much of the deep learning research in radiology have focused on mammograms, chest radiographs, and chest CTs due the availability of publicly available datasets
  • The goal of our FELIX Project is to use deep learning for automatic detection of PDAC from abdominal CTs
  • Automatic detection of PDAC is a difficult task, which can divided into more manageable tasks:
    • Automatic segmentation of abdominal organs
    • Classification of normal vs. abnormal pancreas
Erickson BJ et al. RadioGraphics 2017;37:505-15. Kohli M et al. AJR 2017;208:754-60. Chu LC et al. J Am Coll Radiol. 2019:16:1338-1342

 

Deep Learning – Our Experience

  • Deep learning requires large amount of high quality input data
  • We have manually segmented abdominal organs and vasculature on dual-phase abdominal CTs from:
    • >750 PDAC patients
    • >700 normal controls without known pancreatic disease
Deep Learning - Our Experience

Park S et al. Diagn Interv Imaging. 2019 [Epub ahead of print]. Yan W et al. Medical Image Analysis. 2019;55:88-102

 

Deep Learning – Our Experience

Our deep learning algorithm can achieve >85% accuracy in segmentation of major abdominal organs

Deep Learning - Our Experience

*Segmentation accuracy as defined by Dice-Sørensen coefficient

Park S et al. Diagn Interv Imaging. 2019 [Epub ahead of print]. Yan W et al. Medical Image Analysis. 2019;55:88-102

 

Deep Learning – Our Experience

Our deep learning algorithm can detect PDAC with >90% sensitivity and >90% specificity with a range of sizes and appearances



Deep Learning - Our Experience

Zhu Z et al. MICCAI 2019. Chu LC et al. J Am Coll Radiol. 2019;16:1338-1342.

 

Deep Learning – Ongoing Improvements

  • Enrich training dataset with small and challenging tumors to improve the performance of the deep network
  • Include other tumors types (e.g. PNETs, cystic neoplasms) and pancreatitis for detection and classification
  • Test our algorithm on datasets from other institutions to ensure its performance with other scanner types and scan protocols

 

Advanced Visualization Techniques

  • CT imaging plays a critical role in triaging patients with PDAC into resectable, borderline resectable, and unresectable categories and guide management decisions
  • Multiplanar and 3D reconstructions are helpful in defining extent of vascular involvement and in determining vascular reconstruction options
  • Advanced visualization becomes even more important in cases of planned minimally invasive surgery, since the field-of-view can be limited
Soloff EV et al. Abdom Radiol (NY). 2018:43(2):301-313. Zaky AM et al. Radiographics. 2017:37(1):93-112. Broucek JR et al. Surg Oncol Clin N Am. 2019:28(2):255-272.

 

Cinematic Rendering (CR)

  • Cinematic rendering (CR) is a recently described rendering technique inspired by quality of computer animation programs
  • It uses a global illumination model that take direct and indirect illumination into account to generate photorealistic images


Chu LC et al. Abdom Radiol (NY). 2018:43(11):3009-3015.

 

Cinematic Rendering (CR)

CR increases depth perception and accentuates texture differences and can potentially improve appreciation of:

Cinematic Rendering (CR)

Chu LC et al. Abdom Radiol (NY). 2018:43(11):3009-3015.

 

Current State-of-the-Art

Current State-of-the-Art

Radiologists rely on visual cues to detection and stage pancreatic tumors.

 

Our Vision of Future of Pancreatic Imaging

Our Vision of Future of Pancreatic Imaging

AI functions as second-reader to differentiate abnormal from normal cases, provide probabilities of most likely diagnosis, and extract imaging biomarkers relevant to patient management.

 

Conclusions

  • State-of-the-art CT imaging has excellent accuracy in detection and staging of pancreatic ductal adenocarcinoma
  • Several emerging techniques have the potential to improve upon pancreatic cancer imaging and provide additional value
  • Artificial intelligence (AI) may function as a second-reader for automatic detection of pancreatic cancer, which can lead to earlier detection and improve patient outcomes

 

Conclusions

  • AI augmented workflow can automatically extract quantifiable imaging biomarkers for tumor classification and patient prognostication
  • These additional information will be available to the radiologists and clinicians to guide management decisions
  • Advanced visualization techniques can improve appreciation of tumor extension and assist in preoperative planning

 

References

  • Aerts HJ et al. Nat Commun. 2014;5:4006.
  • Al-Hawary MM et al. Radiology. 2014;270:248-60.
  • Attiyeh MA et al. Ann Surg Oncol. 2018;25:1034-1042.
  • Attiyeh MA et al. HPB (Oxford). 2019;21(2):212-218.
  • Bilici A. World J Gastroenterol. 2014;20(31):10802-10812.
  • Brennan MF et al. Ann Surg. 2004;240:293-298.
  • Broucek JR et al. Surg Oncol Clin N Am. 2019:28(2):255-272.
  • Cassinotto C et al. Eur J Radiol. 2017;90:152-158.
  • Chakraborty J et al. PLoS One. 2017;12(12):e0188022.
  • Chakraborty J et al. Med Phys. 2018;45(11):5019-5029.
  • Chu LC et al. Cancer J. 2017;23(6):333.342.
  • Chu LC et al. Abdom Radiol (NY). 2018:43(11):3009-3015.
  • Chu LC et al. J Am Coll Radiol. 2019;16:1338-1342.
  • Eilaghi A et al. BMC Medical Imaging. 2017;17(1):38.
  • Erickson BJ et al. RadioGraphics 2017;37:505-15.
  • Geras KJ et al. Radiology. 2019 [Epub ahead of print]
  • Gillies RJ et al. Radiology. 2016;278(2):563-577.
  • Gonoi W et al. Eur Radiol. 2017. 27(12):4941-4950.
  • Hanania AN et al. Oncotarget. 2016;7(52):85776-85784.
  • Kohli M et al. AJR 2017;208:754-60.
  • Kumar V et al. Magn Reson Imaging. 2012;30(9):4-48.
  • Laffan TA et al. AJR. 2008;191:802-807.
  • Park S et al. Diagn Interv Imaging. 2019 [Epub ahead of print].
  • Pawlik TM et al. Ann Surg Oncol. 2008. 15(8):2081-8.
  • Permuth JB et al. Oncotarget. 2016;7(52):85785-85797.
  • Soloff EV et al. Abdom Radiol (NY). 2018:43(2):301-313.
  • Yan W et al. Medical Image Analysis. 2019;55:88-102
  • Yun G et al. Scientific Reports. 2018;8:7226.
  • Zanini N et al. Pancreatology. 2015;15:417-422.
  • Zaky AM et al. RadioGraphics. 2017;37:93-112.
  • Zhu Z et al. MICCAI 2019.
  • https://seer.cancer.gov.
© 2020 Elliot K. Fishman, MD, FACR
All Rights Reserved.
www.CTISUS.com
CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning CTisus CT Scanning