CN-121999292-A - Colorectal early cancer infiltration depth prediction model construction method based on mixed cutting enhancement

CN121999292ACN 121999292 ACN121999292 ACN 121999292ACN-121999292-A

Abstract

The invention discloses a colorectal early cancer infiltration depth prediction model construction method based on mixed cutting enhancement, which comprises the steps of preprocessing historical white light endoscopic images, increasing sample size on the premise of highlighting focus areas to obtain a preprocessed sample set, obtaining focus areas ROI in the preprocessed sample set by a doctor and/or adopting a feature recognition algorithm, marking shallow infiltration type labels and deep infiltration type labels on the focus areas ROI, enhancing data diversity and model generalization by CutMix, obtaining enhanced label data, taking the historical white light endoscopic images, the enhanced label data and the type labels as training data, and taking the enhancement training of a basic model by means of a mixed cutting enhancement strategy into consideration, so that the colorectal early cancer infiltration depth prediction model obtained by training predicts shallow infiltration/deep infiltration types after being input into the white light endoscopic images, and objective judgment of colorectal early cancer infiltration depth under a conventional white light endoscope is realized.

Inventors

ZHONG YUNSHI
SUN JINGHAN
WEI DONG
LI BING
ZHOU TAILIN
CAI SHILUN
QI ZHIPENG
LIU JINGYI
Dong Yuelun
YU YUE
WU JIANRONG

Assignees

复旦大学附属中山医院

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (3)

1. The method for constructing the colorectal early cancer infiltration depth prediction model based on mixed cutting enhancement is characterized by comprising the following steps of: the image preprocessing module is used for preprocessing the historical white light endoscopic image, and increasing the sample size on the premise of highlighting the focus area to obtain a preprocessed sample set; the focus area positioning module is used for acquiring focus areas ROI in the sample set after pretreatment by a doctor and/or adopting a feature recognition algorithm, labeling focus areas ROI with 'shallow infiltration' and 'deep infiltration' category labels, enhancing and improving data diversity and model generalization by CutMix, and obtaining label data after enhancement; The model construction module is used for taking the historical white light endoscopic image, the reinforced label data and the class label as training data, and carrying out reinforced training on the basic model by means of a mixed cutting reinforcement strategy, so that the colorectal early cancer infiltration depth prediction model obtained by training predicts shallow infiltration/deep infiltration class after the white light endoscopic image is input.
2. The method for constructing a colorectal early cancer infiltration depth prediction model based on mixed-cut enhancement according to claim 1, wherein the preprocessing comprises normalization, color dithering, clipping, rotation and mirroring.
3. The method for constructing the colorectal early cancer infiltration depth prediction model based on mixed cut enhancement according to claim 1, wherein the colorectal early cancer infiltration depth prediction model is obtained by training a basic model through a minimum cross entropy loss function, and a minimum cross entropy loss function formula is as follows: in the formula, In order to predict the shallow/deep wetting class, Is a category label.

Description

Colorectal early cancer infiltration depth prediction model construction method based on mixed cutting enhancement Technical Field The invention belongs to the technical field of prediction model construction, and particularly relates to a colorectal early cancer infiltration depth prediction model construction method based on mixed cutting enhancement. Background Colonoscopy is an important tool in the screening and diagnosis of colorectal cancer. Polyps found during endoscopy require pathological analysis to determine their nature, particularly to determine the presence of cancerous lesions and the depth of infiltration of cancerous cells. Depending on the depth of infiltration, and in particular whether or not a deep submucosal infiltration is reached, it is critical to formulate a subsequent treatment regimen (endoscopic resection or surgery). With the continuous development of artificial intelligence technology, a deep learning model has been gradually applied to digestive endoscopy image analysis for colorectal polyp detection, benign and malignant discrimination and pathological risk classification. The partial model shows the diagnosis level approaching or even exceeding the diagnosis level of the senior doctor under the experimental condition, and shows the potential of artificial intelligence in improving the diagnosis efficiency and reducing the subjective dependence. However, from the clinical application point of view, the prior art still has obvious defects that most researches only stay at the focus recognition and benign and malignant differentiation level, and the key index directly determining the treatment strategy, namely the infiltration depth, lacks an effective solution. The current clinic mainly depends on specific imaging modes such as EUS, NBI and the like and the experience of doctors to judge the focus depth, but the methods are limited by equipment conditions and operator level, and have the problems of long learning curve, low popularity, strong subjectivity and the like. In addition, the partial deep learning system has higher requirements on image quality, and is not beneficial to wide popularization in the conventional endoscope process and basic medical institutions. In current clinical practice, there is still a lack of an auxiliary tool that can rapidly, stably and objectively determine whether colorectal early cancer reaches deep infiltration based on conventional white light endoscopic images. In practice, the core problem that endoscopists often need to solve is "whether the focus is suitable for endoscopic resection", but this determination has great difficulty, directly affecting the accuracy of treatment decisions, and possibly even causing excessive treatment or delayed treatment. Therefore, development of an artificial intelligent auxiliary tool which can be directly embedded into a conventional inspection process without depending on additional equipment or special dyeing is needed to improve diagnosis and treatment accuracy and safety. Disclosure of Invention Aiming at the lack of a tool for objectively judging the infiltration depth of colorectal early cancer under a conventional white light endoscope in the prior art, the invention provides an artificial intelligence solution based on Swin transducer and CutMix enhancement, and aims to provide an artificial intelligence auxiliary judgment method which is low in threshold, high in efficiency and popular, and directly serves for clinical core decision of whether endoscope excision can be performed. In order to achieve the above object, the present invention provides a method for constructing a colorectal early cancer infiltration depth prediction model based on mixed-cut enhancement, comprising the following steps: the image preprocessing module is used for preprocessing the historical white light endoscopic image, and increasing the sample size on the premise of highlighting the focus area to obtain a preprocessed sample set; the focus area positioning module is used for acquiring focus areas ROI in the sample set after pretreatment by a doctor and/or adopting a feature recognition algorithm, labeling focus areas ROI with 'shallow infiltration' and 'deep infiltration' category labels, enhancing and improving data diversity and model generalization by CutMix, and obtaining label data after enhancement; the model construction module is used for taking the history white light endoscopic image, the reinforced label data and the class label as training data, and carrying out reinforced training on the basic model by means of a mixed cutting reinforcement strategy, so that the colorectal early cancer infiltration depth prediction model obtained by training predicts shallow infiltration/deep infiltration class after the white light endoscopic image is input. Preferably, the preprocessing includes normalization, color dithering, clipping, rotation, mirroring. Preferably, the colorectal early cancer infiltration depth p