Strategies to improve performance of convolutional neural network on histopathological images classification

Convolutional Neural Network (CNN) has been widely used in medical image processing. Histopathology is one of modality or images for a pathologist to analyze the status of cancer. The unstructured pattern of this image cause the problem, tend to miss identification or takes more time to analyze by the pathologist. Besides that, Deep learning training generally requires powerful hardware resources to improve performance during the training. Therefore, to address these problems, we propose two main activities in this study; to accelerate training time and to enhance the histopathology dataset. We train our CNN on three similar GPU specification (GTX-1080) as an alternative to become training time is faster. Mean-shift filter is one of the low-pass filter technique. We use this to handle unstructured pattern on histopathology images to enhance this dataset. The performance of all three GPUs is presented during the training process with 500 epochs measure by the speedup. Meanwhile, the performance of model testing is carried out with several batch-size selection scenarios from 32,64,128 and 256. The use of mean-shift can improve convergence during training in 128 batch-size become faster.