Multi-Order Transfer Learning For Pathologic Diagnosis Of Pulmonary Nodule Malignancy

Precise diagnosis of pulmonary nodules can be very crucial as pulmonary nodules are often the common manifestation of early lung cancers. In this study, we investigate the multi-order transfer learning for the assessments of pulmonary nodules to leverage the classification performance of nodules with pathologic confirmation in the condition of small samples. The experiments show that the 3rd order transfer with the source tasks of texture, diameter and lobulation can achieve the best performance (Acc=0.8194, AUC=0.7533) among all 10 orders transfer learning in the pathologic diagnosis (golden standard) of nodule malignancy, which shows a higher performance than the state-of-the art methods and even outperforms radiologists' performance (Acc=0.7241, AUC=0.76) in terms of Accuracy. This multi-order transfer learning is shown to be effective in the pathologic diagnosis of pulmonary nodule malignancy with simply need only 30% semantic tasks as source tasks.