This work presents comprehensive results to detect in the early stage the pancreatic neuroendocrine tumors (PNETs), a group of endocrine tumors arising in the pancreas, which are the second common type of pancreatic cancer, by checking the abdominal CT scans. To the best of our knowledge, this task has not been studied before as a computational task. To provide radiologists with tumor locations, we adopt a segmentation framework to classify CT volumes by checking if at least a sufficient number of voxels is segmented as tumors. To quantitatively analyze our method, we collect and voxelwisely label a new abdominal CT dataset containing $376$ cases with both arterial and venous phases available for each case, in which $228$ cases were diagnosed with PNETs while the remaining $148$ cases are normal, which is currently the largest dataset for PNETs to the best of our knowledge. In order to incorporate rich knowledge of radiologists to our framework, we annotate dilated pancreatic duct as well, which is regarded as the sign of high risk for pancreatic cancer. Quantitatively, our approach outperforms state-of-the-art segmentation networks and achieves a sensitivity of $89.47\%$ at a specificity of $81.08\%$, which indicates a potential direction to achieve a clinical impact related to cancer diagnosis by earlier tumor detection.