Deep learning for pancreatic cancer detection: current challenges and future strategies

Abstract

In The Lancet Digital Health, Kao-Lang Liu and colleagues1Liu K-L Wu T Chen P-T et al.Deep learning to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue: a retrospective study with cross-racial external validation.Lancet Digital Health. 2020; 2: e303-e313Summary Full Text Full Text PDF Scopus (16) Google Scholar describe the applications of a convolutional neural network (CNN) in distinguishing CT images of pancreatic cancer tissue from non-cancerous pancreatic tissue.1Liu K-L Wu T Chen P-T et al.Deep learning to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue: a retrospective study with cross-racial external validation.Lancet Digital Health. 2020; 2: e303-e313Summary Full Text Full Text PDF Scopus (16) Google Scholar Contrast-enhanced CT images of 370 patients with pancreatic cancer and 320 controls from a Taiwanese centre were manually segmented and CNN was trained to classify image patches as cancerous or non-cancerous. Performance of CNN was compared with radiology reports in the local test sets. Similar to results from previous studies on this topic,2Chu LC Park S Kawamoto S et al.Utility of CT radiomics features in differentiation of pancreatic ductal adenocarcinoma from normal pancreatic tissue.AJR Am J Roentgenol. 2019; 213: 349-357Crossref PubMed Scopus (46) Google Scholar, 3Chu LC Park S Kawamoto S et al.Application of deep learning to pancreatic cancer detection: lessons learned from our initial experience.J Am Coll Radiol. 2019; 16: 1338-1342Summary Full Text Full Text PDF PubMed Scopus (23) Google Scholar CNN achieved a remarkable performance, with accuracy of 0·986–0·989 in the local test dataset. CNN achieved higher sensitivity than that of radiologists in the local test sets (0·983 vs 0·929; p=0·014). The three tumours that were missed by the CNN were 1·1–1·2 cm in size, of which two were correctly classified by radiologists. Impressively, CNN was able to correctly classify 11 of the 12 tumours that were missed by radiologists, which were 1·0–3·3 cm in size.1Liu K-L Wu T Chen P-T et al.Deep learning to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue: a retrospective study with cross-racial external validation.Lancet Digital Health. 2020; 2: e303-e313Summary Full Text Full Text PDF Scopus (16) Google Scholar These promising results show that CNN as a second reader can reduce misdiagnosis of pancreatic cancer and possibly lead to improved patient outcomes. Whereas previous publications relied on data from a single institution,2Chu LC Park S Kawamoto S et al.Utility of CT radiomics features in differentiation of pancreatic ductal adenocarcinoma from normal pancreatic tissue.AJR Am J Roentgenol. 2019; 213: 349-357Crossref PubMed Scopus (46) Google Scholar, 3Chu LC Park S Kawamoto S et al.Application of deep learning to pancreatic cancer detection: lessons learned from our initial experience.J Am Coll Radiol. 2019; 16: 1338-1342Summary Full Text Full Text PDF PubMed Scopus (23) Google Scholar Liu and colleagues also externally validated data with publicly available datasets from the USA, with an accuracy of 0·832. Importantly, external datasets of individuals with normal pancreas and pancreatic cancer patients originated from different sources: 82 patients with healthy pancreas from the Cancer Imaging Archive dataset and 281 patients with pancreatic cancer from the Medical Segmentation Decathlon dataset from Memorial Sloan Kettering Cancer Center, New York, NY, USA. Ideally, all patients from the external data should be from the same institution to ensure that the classification task is not affected by technical differences between the two institutions. The authors showed that a significant positive correlation existed between tumour size and CNN performance on both the internal and external datasets, which suggested that the CNN detected actual differences in pancreatic imaging features, rather than technical differences. These preliminary experiences in applying a CNN to detect pancreatic cancer show tremendous promise, but more external validation with datasets from multiple institutions is needed. The rate-limiting step lies in the curation of large, high-quality, annotated datasets that are publicly available, which will allow researchers to compare the performance of CNN models.4Langlotz CP Allen B Erickson BJ et al.A roadmap for foundational research on artificial intelligence in medical imaging: from the 2018 NIH/RSNA/ACR/The Academy Workshop.Radiology. 2019; 291: 781-791Crossref PubMed Scopus (92) Google Scholar However, curation of imaging data is an expensive and labour-intensive process that requires expertise in imaging anatomy and pathology, which limits the ability for crowdsourcing.5Park S Chu LC Fishman EK et al.Annotated normal CT data of the abdomen for deep learning: challenges and strategies for implementation.Diagn Interv Imaging. 2020; 101: 35-44Crossref Scopus (19) Google Scholar, 6Prevedello LM Halabi SS Shih G et al.Challenges related to artificial intelligence research in medical imaging and the importance of image analysis competitions.Radiology: Artificial Intelligence. 2019; 1e180031Crossref Scopus (28) Google Scholar Additionally, logistical and ethical concerns exist regarding data ownership and the protection of patient privacy that might hinder the development and sharing of public imaging datasets.6Prevedello LM Halabi SS Shih G et al.Challenges related to artificial intelligence research in medical imaging and the importance of image analysis competitions.Radiology: Artificial Intelligence. 2019; 1e180031Crossref Scopus (28) Google Scholar, 7Larson DB Magnus DC Lungren MP Shah NH Langlotz CP Ethics of using and sharing clinical imaging data for artificial intelligence: a proposed framework.Radiology. 2020; (published online March 24.)DOI:10.1148/radiol.2020192536Crossref Scopus (19) Google Scholar To streamline the process, the Early Detection Research Network of the National Cancer Institute is in the preliminary stages of organising a centralised repository of pancreatic cancer imaging data. Liu and colleagues showed that a CNN was able to achieve similar or even superior performance compared with that of radiologists. The goal for pancreatic cancer detection will be identifying pancreatic cancer before the subtle visual changes are apparent to a radiologist. One of the challenges in achieving this goal is the paucity of training data with these early subtle pancreatic cancers, because average-risk patients are not routinely screened for pancreatic cancer. The solution might be found with surveillance programmes in which patients at high risk of developing pancreatic cancer, based on family history or germline mutation, undergo serial imaging studies.8Goggins M Overbeek KA Brand R et al.Management of patients with increased risk for familial pancreatic cancer: updated recommendations from the International Cancer of the Pancreas Screening (CAPS) Consortium.Gut. 2020; 69: 7-17Crossref Scopus (91) Google Scholar In relation to the patients who eventually developed pancreatic cancer, the imaging exams before the diagnosis was made by the radiologist (pre-diagnostic exams) can be used retrospectively to train the CNN to recognise subtle imaging features that might signal the presence of an early stage pancreatic cancer. Another potential approach is to harness the longitudinal clinical and imaging data in large private or nationalised health systems, which might be able to provide such pre-diagnostic exams for training and testing. Finally, Liu and colleagues' study, as with previous studies on pancreatic cancer detection,2Chu LC Park S Kawamoto S et al.Utility of CT radiomics features in differentiation of pancreatic ductal adenocarcinoma from normal pancreatic tissue.AJR Am J Roentgenol. 2019; 213: 349-357Crossref PubMed Scopus (46) Google Scholar, 3Chu LC Park S Kawamoto S et al.Application of deep learning to pancreatic cancer detection: lessons learned from our initial experience.J Am Coll Radiol. 2019; 16: 1338-1342Summary Full Text Full Text PDF PubMed Scopus (23) Google Scholar has been presented as including artificial binary constructs: cancer versus normal. However, radiologists encounter various types of pancreatic tumours in clinical practice, and non-neoplastic conditions such as pancreatitis can simulate the appearance of pancreatic tumors.9Haj-Mirzaian A Kawamoto S Zaheer A Hruban RH Fishman EK Chu LC Pitfalls in the MDCT of pancreatic cancer: strategies for minimizing errors.Abdom Radiol. 2020 Feb; 45: 457-478Crossref Scopus (4) Google Scholar, 10Park S Chu LC Hruban RH et al.Differentiating autoimmune pancreatitis from pancreatic ductal adenocarcinoma with CT radiomics features.Diagn Interv Imaging. 2020; (pubished online April 8.)DOI: 10.1016/j.diii.2020.03.002Crossref Scopus (22) Google Scholar Additional training will be necessary to distinguish pancreatic cancer from these other pathologies to improve the clinical utility of the model. Artificial intelligence will certainly revolutionise the practice of radiology within the next decade. The collaboration of radiologists, computer scientists, and cancer researchers and the sharing of domain expertise will be essential in powering this revolution. We receive research grant support from The Lustgarten Foundation and The Emerson Collective. We thank senior science editor Edmund Weisberg for his editorial assistance. Deep learning to distinguish pancreatic cancer tissue from non-cancerous pancreatic tissue: a retrospective study with cross-racial external validationCNN could accurately distinguish pancreatic cancer on CT, with acceptable generalisability to images of patients from various races and ethnicities. CNN could supplement radiologist interpretation. Full-Text PDF Open Access

Topics & Concepts