Abstract
Automatic pattern recognition using deep learning techniques has become increasingly important. Unfortunately, due to limited system memory, general preprocessing methods for high-resolution images in the spatial domain can lose important data information such as high-frequency information and the region of interest. To overcome these limitations, we propose an image segmentation approach in the compressed domain based on principal component analysis (PCA) and discrete wavelet transform (DWT). After inference for each tile using neural networks, a whole prediction image was reconstructed by wavelet weighted ensemble (WWE) based on inverse discrete wavelet transform (IDWT). The training and validation were performed using 351 colorectal biopsy specimens, which were pathologically confirmed by two pathologists. For 39 test datasets, the average Dice score, the pixel accuracy, and the Jaccard score were 0.804 ± 0.125, 0.957 ± 0.025, and 0.690 ± 0.174, respectively. We can train the networks for the high-resolution image with the large region of interest compared to the result in the low-resolution and the small region of interest in the spatial domain. The average Dice score, pixel accuracy, and Jaccard score are significantly increased by 2.7%, 0.9%, and 2.7%, respectively. We believe that our approach has great potential for accurate diagnosis.
Highlights
Automatic pattern recognition using deep learning techniques has become increasingly important
For the model using information loss data (Small ROI, low resolution), the average Dice, Acc, and Jaccard score (Jac) results decreased by 1.1%, 0.2%, and 1.2% for the small ROI data (Tile size: 256 by 256, 20× magnification) and 4%, 1%, and 4.4% for the low resolution data (Tile size: 512 by 512, 10× magnification) respectively, compared to those of the model using standard data (Tile size: 512 by 512, 20× magnification)
For the model using compressed data, the average Dice, Acc, and Jac results for the LL sub-band increased by 4%, 0.6%, and 4.3%, respectively, compared to those of the model using low resolution data in spatial domain whose magnification equal to LL sub-band’s magnification
Summary
Automatic pattern recognition using deep learning techniques has become increasingly important. Due to limited system memory, general preprocessing methods for high-resolution images in the spatial domain can lose important data information such as high-frequency information and the region of interest. To overcome these limitations, we propose an image segmentation approach in the compressed domain based on principal component analysis (PCA) and discrete wavelet transform (DWT). Decimation can cause a loss of high-frequency information, resulting in low resolution due to the reduced signal bandwidth[24,25] As another widely used method, cropping extracts the wanted areas from whole slide images (WSIs) into tiles.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have