Holistic Analysis of Abdominal CT for Predicting the Grade of Dysplasia of Pancreatic Lesions.

Diagnosis of various pancreatic lesions in CT images is a challenging task owing to a significant overlap in their imaging appearance. An accurate diagnosis of pancreatic lesions and the assessment of their malignant progression, or the grade of dysplasia, is crucial for optimal patient management. Typically, the grade of dysplasia is confirmed histologically via biopsy, yet certain radiological findings, including extrapancreatic, can serve as diagnostic clues of the disease progression. This work introduces a novel method of transforming intermediate activations for processing intact imaging data of varying sizes with convnets with linear layers. Our method allows to efficiently leverage the 3D information of the entire abdominal CT scan to acquire a holistic picture of all radiological findings for an improved and more precise classification of pancreatic lesions. Our model outperforms current state-of-the-art methods in classifying four most common lesion types (by 2.92%), while additionally diagnosing the grade of dysplasia. We conduct a set of experiments to illustrate the effects of a holistic CT analysis and the auxiliary diagnostic data on the accuracy of the final diagnosis.