Abstract: Aiming at the problems of small data set,uneven distribution of benign and malignant samples,and low automatic re-cognition accuracy of breast cancer pathological tissue image data,a lightweight pathological tissue image classification model with reasonable depth and width is designed,which is suitable for small data sets.Based on the traditional data enhancement methods such as image rotation and distortion,the random non-repeated cutting method is used to balance the number of benign and malignant samples and expand the data set.For the samples that are difficult to cluster in the training set,the concept of “weak feature”,“weak feature” sample extraction algorithm and adaptive adjustment,secondary training algorithm are proposed to improve the model training.Under the condition of the same parameter setting and running environment,eight groups of comparative experiments are carried out,and the accuracy,sensitivity and specificity of the model can reach more than 97%.The experimental results show that the performance of the model designed in this paper is stable,and it has good tolerance and adaptability for small data sets and unbalanced data sets.

Key words: Adaptive small data sets, Breast cancer pathological tissue images, Convolutional neural networks, Deep learning, Deep separable convolution, Weak features