Parallel 'same' and 'valid' convolutional block and input-collaboration strategy for histopathological image classification

Histopathological image classification is of great importance in pathological diagnosis, such as tumor grading and cancer type identification. However, the traditional pathological examination is timeconsuming and requires the subjective judgments and rich experience of pathologists. In order to alleviate these problems and provide quantitative analysis results, this paper proposes a parallel 'same' and 'valid' convolutional block (PSV-CB) and an input-collaboration strategy for performing histopathological image classification. The core of PSV-CB is to employ different convolutional operations for learning hidden representations of each input respectively and then correspondingly integrate them together to highlight interesting contents, where an operational flow is constructed via multiple 'same' convolutions and followed by a max-pooling, which can be considered as a 'hard' feature coding. Another one is established using step-by-step 'valid' convolutions that consider feature extraction and down-sampling simultaneously, which can be regarded as a 'soft' feature coding. Therefore, the parallel 'same' and 'valid' convolutional neural network (PSV-ConvNet) is constructed using stacked PSV-CB according to the specific task. To compensate the loss of spatial information, we introduce an input-collaborative PSV-ConvNet (IC-PSV-ConvNet) that adds grayscale version of original inputs into the outputs of each PSV-CB for better fusing global information and learned features. The proposed IC-PSV-ConvNet is evaluated on three histopathological image datasets (lymphoma, breast cancer, and liver cancer) with satisfactory results. Our comparative experiments demonstrate that the proposed IC-PSV-ConvNet can achieve more accurate classification results compared to other existing ConvNets. (c) 2022 Elsevier B.V. All rights reserved.