Gastric cancer digital pathological section detection method based on improved VGG16 network

Abstract

The invention discloses a stomach cancer digital pathological section detection method based on an improved VGG16 network, and the method comprises the following steps: extracting an ROI region based on a rapid Otsu algorithm, and generating an initial annotation mask graph based on an xml annotation file; establishing a model based on the VGG16 network, retaining a convolutional layer and a pooling layer of the VGG16 network, combining two characteristic maps with the same size in a fully connected layer of the VGG16 network to obtain a new map, inputting the combined new map into a classification function Softmax to classify each pixel, and obtaining output; according to the method, on the premise that the prediction effect is maintained, the parameter quantity and the calculation quantity are greatly reduced from the aspects of data processing, model training, image prediction and the like, fewer resources are used, and sketching of the gastric cancer focus area is completed.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a method for detecting digital pathological slices of gastric cancer based on an improved VGG16 network. Background technique [0002] Many scholars have applied traditional machine learning algorithms to WSI and achieved certain results. In recent years, deep learning has also been widely used in pathological image research. At present, models based on CNN and UNet have realized the detection of digital pathological slices of gastric cancer, but these methods still have some problems: [0003] (1) The data set production time is long. The pixel level of the original pathological section is usually 10 5 ×10 5 It takes dozens of hours to process a WSI using the general Otsu algorithm, and it often takes several days or even a dozen days to process the entire data set, which is time-consuming and labor-intensive and increases research costs; [0004] (2) The model comp...

AI Technical Summary

Problems solved by technology

The pixel level of raw pathology slides is usually 10 5 ×10 5 It takes dozens of hours to process a WSI with the general Otsu algorithm, and it often takes several days or even ten days to process the entire data set, which is time-consuming and labor-intensive, and increases research costs;

[0004] (2) High model complexity

The convolution kernel used by the general model is relatively large and contains a huge amount of parameters, which requires a lot of hardware resources and time costs during training, and the efficiency of prediction is not high;

[0005] (3) Time-consuming model training

WSI is usually sized in terms of 10 5 ×10 5 is the unit, usually needs to be cut to 10 when inputting the model 2 ×10 2 Therefore, a WSI will cut at least hundreds of thousands of patches, and the data set made from dozens of original pathological slices will contain millions or even tens of millions of patches, and multiple rounds of iterations are required during training. The whole training is very time-consuming, consumes huge computing resources, and is not conducive to adjusting parameters during training.

Images