Fusion strategies for deep convolutional neural network representations in histopathological image classification

The enormous rise in processing power and the refinement of analysis algorithms over the past decade have resulted in the creation of several ways of computer-aided medical data analysis. We investigated the merits of combining a set of image classification systems incorporating 16 pre-trained convolutional neural networks with classifiers fusion strategy. We suggested a fusion paradigm that works on the posterior probability of different classifiers based on deep representations. For this purpose, motivated by the scarcity of medical images for training deep networks, a transfer learning paradigm was adopted where different deep convolutional networks (pre-trained on non-medical images) were used as mechanisms for feature extraction. Two well-known datasets including the animal diagnostics laboratory (ADL) and the breast cancer histopathological image classification (BreakHis) datasets were used. In the ADL dataset, the best result from the combination of classifiers trained on two pre-trained networks was 100% accuracy belong to the lung data. In the BreaKHis dataset, the suggested technique was investigated on both the image and patient levels which got the highest percentages of image classification rates compared to prior methods. This approach at 400 magnification achieved 92.7% at the image level, respectively, and 93.2% at the patient level. The Fisher's least significant difference post hoc test results showed a significant difference in accuracy parameters across the three classifiers reported in this research in the ADL and BreaKHis dataset with a p value < 0.001. The proposed classifier fusion strategy meets the best-reported performance for classifying histopathological images on two different histopathological datasets, ADL, and BreaKHis.