An Efficient Neural Network based on Early Compression of Sparse CT Slice Images

Recently, research on diagnosing diseases through artificial intelligence has been conducted in various medical fields, including Thyroid-associated ophthalmopathy. We introduce a computationally efficient CNN architecture, which is optimized for CT images and designed especially for mobile devices with very limited computing power. The proposed architecture utilizes three operations, pointwise convolution, depth-wise separable convolution and channel shuffle, to reduce computation cost for handling a series of CT image slices for a patient. On CT images, the proposed model achieves ∼ 3.5 × actual speedup over ShuffleNet-v2 without degenerating prediction accuracy.