Cancer prediction from few amounts of histology samples through self-attention based multi-routines cross-domains network.

Objective: Rapid and efficient analysis of cancer has become a focus of research. Artificial intelligence can use histopathological data to quickly determine the cancer situation, but still faces challenges. For example, the convolutional network is limited by the local receptive field, human histopathological information is precious and difficult to be collected in large quantities, and cross-domain data is hard to be used to learn histopathological features. In order to alleviate the above questions, we design a novel network, Self-attention based multi-routines cross-domains network (SMC-Net).

Approach: Feature analysis module and decoupling analysis module designed are the core of the SMC-Net. The feature analysis module base on multi-subspace self-attention mechanism with pathological feature channel embedding. It in charge of learning the interdependence between pathological features to alleviate the problem that the classical convolution model is difficult to learn the impact of joint features on pathological examination results. The decoupling analysis module base on the designed multi-channel and multi-discriminator architecture. Its function is to decouple the features related to the target task in cross-domain samples so that the model has cross-domain learning ability.

Main Results: To evaluate the performance of the model more objectively, three datasets are used. Compared with other popular methods, our model achieves better performance without performance imbalance. In this work, a novel network is design. It can use domain-independent data to assist in the learning of target tasks, and can achieve acceptable histopathological diagnosis results even in the absence of data.

Significance: The proposed method has higher clinical embedding potential and provides a viewpoint for the combination of deep learning and histopathological examination.