Preliminary assessment for the development of CADe system for brain tumor in MRI images utilizing transfer learning in Xception model

Brain tumor is one of the most detrimental diseases. Magnetic resonance imaging (MRI) has been used for the diagnosis of the brain tumor; however, manual tumor diagnosis in MRIs is a time-consuming process, necessitating a development of automatic brain tumor diagnosis system. This study applied a novel architecture, named Xception, which enabled both high performance and reduced size and computational cost of convolutional neural networks (CNNs) using depthwise separable convolution to develop high-performance computer aided diagnosis (CADe) system for brain tumor detection from MRI. Preliminary assessment for the Xception model utilizing transfer learning demonstrated good performance with high accuracy and prediction probability. Interestingly prediction probabilities were different when different layers were relearned. The prediction probability values were highest in both normal and tumor cases when the 109
^th
layer and after (ie, the exit flow) were relearned, indicating the importance of re-learning for the Exit flow. The study results suggest that a high-performance CADe system for brain tumor in MRIs could be developed with relatively cheap small-scale learning utilizing transfer learning in the Xception model.