The diagnosis of gastric atrophy is highly subjective. We aim to establish a model of gastric atrophy based on pathological characteristics to improve the consistency of diagnosis.

HE-stained pathological sections of gastric biopsies were retrospectively collected, and CellProfiler software was used to perform image segmentation and feature extraction on 10 representative images of each sample.

Building a gastric atrophy diagnostic model using LASSO feature selection and different machine learning (ML) algorithms

After normalizing the data, the Pearson correlation coefficient between features was calculated. If the correlation coefficient between features is greater than 0.9, only one feature is retained

Divide the data into training set and validation set in a ratio of 8:2

We selected 289 pathological slides, of which 169 were used for model training, 42 were used for testing, and 78 pathological slides from different years were used for external validation.

9 machine learning models

Two professional pathologists independently diagnosed 289 slices, and the diagnosis was consistent, with a Kappa value of 0.68 (95% CI: 0.60 ~ 0.77, P < 0.001)

Disagreeing slides were jointly assessed by two experts. Finally, the two pathologists reached consensus on all discrepant slides through joint discussion

289 gastric biopsy specimens were selected for training, testing and external validation

Extract 464 pathological features, filter out 10 features through LASSO, and establish a diagnostic model for moderate to severe atrophy.

324 features related to nuclear granularity, texture, size, shape and pixel intensity distribution

124 features related to image quality, intensity, co-localization and correlation between intensities

The AUC of various machine learning algorithms is 0.835 ~ 1.000 in the training set, 0.786 ~ 0.949 in the test set, and 0.689 ~ 0.818 in the external validation set

The LR model has the highest AUC value, 0.900 for the training set, 0.901 for the test set, and 0.818 for the external validation set.

According to the Youden index, 0.47 is selected as the cutoff value, and the LR model is converted into a pathological score (PS) of gastric atrophy, including high PS and low PS.

In univariate logistic regression analysis, gender, age and PS were related to GC

Subsequently, multivariate logistic regression analysis was performed, adjusting for gender and age. The results showed that PS is an independent risk factor for GC (OR = 5.70)

The atrophy pathology score based on the LR model was associated with endoscopic atrophy grade (Z = -2.478, P = 0.013) and gastric cancer (OR = 5.70, P < 0.001)

ML model based on pathological features improves diagnostic consistency of gastric atrophy and correlates with endoscopic atrophy grading and gastric cancer

Effectiveness of machine learning in diagnosing moderate to severe gastric atrophy

Comparison of our model with previous models

Single-factor and multi-factor logistic regression analysis of GC

The entire study design workflow

Pathomic feature selection for moderate to severe gastric atrophy based on LASSO algorithm

Histogram of AUC values ​​of different machine learning models for diagnosing moderate to severe gastric atrophy in the training set, test set and external validation set

The efficacy of LR model on moderate to severe gastric atrophy

Performance of logistic regression model with 1000 bootstraps

Comprehensive list of quantitative features extracted by gastric biopsy image processing pipeline

After removing 30 irrelevant or abnormal features, 434 remained

Select the representative area with the best staining effect through Qupath software

Correlation between PS and endoscopic atrophy grade (Kimura Takemoto classification)

The correlation between PS and endoscopic atrophy grade (Kimura Takemoto classification) was explored

The results showed that PS was also correlated with endoscopic atrophy grade (Z = -2.478, P = 0.013)