Background: The diagnostic performance of CT for pancreatic cancer is interpreter-dependent, and approximately 40% of tumours smaller than 2 cm evade detection. Convolutional neural networks (CNNs) have shown promise in image analysis, but the networks’ potential for pancreatic cancer detection and diagnosis is unclear. We aimed to investigate whether CNN could distinguish individuals with and without pancreatic cancer on CT, compared with radiologist interpretation.
Methods: In this retrospective, diagnostic study, contrast-enhanced CT images of 370 patients with pancreatic cancer and 320 controls from a Taiwanese centre were manually labelled and randomly divided for training and validation (295 patients with pancreatic cancer and 256 controls) and testing (75 patients with pancreatic cancer and 64 controls; local test set 1). Images were preprocessed into patches, and a CNN was trained to classify patches as cancerous or non-cancerous. Individuals were classified as with or without pancreatic cancer on the basis of the proportion of patches diagnosed as cancerous by the CNN, using a cutoff determined using the training and validation set. The CNN was further tested with another local test set (101 patients with pancreatic cancers and 88 controls; local test set 2) and a US dataset (281 pancreatic cancers and 82 controls). Radiologist reports of pancreatic cancer images in the local test sets were retrieved for comparison.
Findings:Between Jan 1, 2006, and Dec 31, 2018, we obtained CT images. In local test set 1, CNN-based analysis had a sensitivity of 0·973, specificity of 1·000, and accuracy of 0·986 (area under the curve [AUC] 0·997 (95% CI 0·992–1·000). In local test set 2, CNN-based analysis had a sensitivity of 0·990, specificity of 0·989, and accuracy of 0·989 (AUC 0·999 [0·998–1·000]). In the US test set, CNN-based analysis had a sensitivity of 0·790, specificity of 0·976, and accuracy of 0·832 (AUC 0·920 [0·891–0·948)]. CNN-based analysis achieved higher sensitivity than radiologists did (0·983 vs 0·929, difference 0·054 [95% CI 0·011–0·098]; p=0·014) in the two local test sets combined. CNN missed three (1·7%) of 176 pancreatic cancers (1·1–1·2 cm). Radiologists missed 12 (7%) of 168 pancreatic cancers (1·0–3·3 cm), of which 11 (92%) were correctly classified using CNN. The sensitivity of CNN for tumours smaller than 2 cm was 92·1% in the local test sets and 63·1% in the US test set.
Interpretation:CNN could accurately distinguish pancreatic cancer on CT, with acceptable generalisability to images of patients from various races and ethnicities. CNN could supplement radiologist interpretation.
Funding:Taiwan Ministry of Science and Technology
Methods: CT images of 70 patients with histologically confirmed pancreatic adenocarcinoma and 70 control subjects with normal pancreas were extracted from archive. Images of 10 patients and 10 controls were randomly selected as test set, and others were used as training set. The pancreas and tumor were manually labeled by a radiologist experienced in pancreatic imaging and served as ground truth. Each image was preprocessed by windowing and normalization before cutting into patches of 50 pixels in length and width. Patches in which cancer occupied >30% of the area were labeled as cancer, whereas patches that did not contain cancer and included pancreas occupying>50% of the area, either from controls or PC patients, were labeled as normal pancreas. A convolution neural network (CNN) was trained using one TITAN V (NVIDIA) to classify patches into cancer or normal using binary cross-entropy as loss function, and its performance was assessed on the test set. For patient-based analysis, normal patches from PC patients were excluded, and patients were classified as having PC if more than 50% of his/her patches were predicted as cancer by CNN.
Results: A total of 2522 patches of PC and 3808 patches of normal pancreas were generated and used for training, and the test set included 533 patches of PC and 949 patches of normal pancreas. For differentiation between PC and normal pancreas, the accuracy of the CNN by patch-based analysis was 77.1%, with an area under the receiver operating characteristic (ROC) curve (AUC) of 0.80 (Table and Figure). In patient-based analysis, the accuracy and AUC were 90% and 0.96, respectively. The computation time for analyzing the whole test set (1482 patches) was less than 5 seconds.
Conclusion: We developed a CNN that could differentiate between CTs of pancreas with PC and without PC with a 90% accuracy in patient-based analysis and 77.1% accuracy in patch-based analysis. This deep-learning model is a potential computer-aided diagnosis tool to facilitate early and accurate diagnosis of PC.