Method for automatically predicting gene expression categories based on cancer CT images

Abstract

The invention discloses a method for automatically predicting a gene expression category based on a cancer CT image. The method comprises the following steps: a) acquiring ROI slices and expanding thenumber by 48 times; b) constructing a neural network based on DenseNet-12 and a spatial pyramid module; c) performing training by using a focusing loss function; and d) comprehensively judging the model prediction to obtain a final prediction result. The data expansion technology adopted by the invention can greatly expand the data volume without changing the property of the CT image. A spatial pyramid pooling module with four dimensions extracts multi-level image features, including global semantics and detail features. Focal-Loss is used for guiding a network to pay more attention to slicesof which effective features are difficult to mine at the tumor edge, namely the head end and the tail end, and a training strategy of which the precision is gradually improved is used, so that accurate and efficient CT image gene mutation prediction is finally realized.

Description

technical field [0001] The invention relates to the fields of image processing, computer vision, deep learning, medical image computing and computer-assisted intervention technology (Medical Image Computing and Computer-Assisted Intervention), specifically a method for automatically detecting gene expression types based on cancer CT images. Background technique [0002] Recent studies at home and abroad have shown that the features extracted from cancer CT images are related to certain gene expression patterns. For example, in 2015, Shinagare et al. verified the association between tumor margins, nodular enhancement, and intratumoral blood vessels and VHL mutations. Karlo et al. proposed in 2014 that PBRM1 and SETD2 gene mutations are mainly found in solid (non-cystic) kidneys. in clear cell carcinoma cases. In the past two years, more and more people have begun to explore it. For example, in 2018, Mohammad et al. used a multi-instance learning CNN network to detect the de...

AI Technical Summary

Problems solved by technology

[0003] However, these existing methods use a large amount of medical data, and it is often difficult to collect specific tumor CT data sets with the gold standard of gene mutation status in actual situations.

Moreover, due to the difference in tumor size, location, and shape, existing methods will resample it to a fixed size for training, which will undoubtedly lose the accuracy of the image and ignore the differences between individual tumors

In addition, the tumor edge (i.e., the axial end and the end) of the CT sequence generally contains less tumor parts, and it is difficult to identify features at these slice levels, and the existing methods have not paid attention to this

Images