Deep learning model for detecting high-grade dysplasia in colorectal adenomas

Objective

Early detection and removal of suspicious polyps during routine colonoscopies play an important role in reducing the risk of colorectal cancer. Patient management and follow-up are determined by the type of polyps removed and the degree of dysplasia present on histological evaluation. Whereas discerning between a benign polyp and a dysplastic one is a trivial task, distinguishing between tubular adenomas with low-grade dysplasia (LGD) and high-grade dysplasia (HGD) is a challenging task. In this study, we trained a deep learning model to distinguish between colorectal adenomas with LGD and HGD.

Design

We retrieved 259 slides of adenomatous polyps taken between January 2011 and October 2024. Slides with HGD were reviewed by a subspecialty-trained GI pathologist. After excluding discordant and duplicate cases, 200 slides remained: 71 (35.5%) with HGD and 129 (64.5%) with LGD. The slides were divided into training (160 slides, 80%) and test (40 slides, 20%) sets. After patch generation and stain normalization, a ResNet34 model (pre-trained on ImageNet) was trained using 5-fold cross-validation. Slide classification was determined by aggregating patch-level predictions.

Results

The model's slide-level prediction accuracy was 95.0%, correctly classifying all but 2 out of 40 slides. The model achieved an area under the receiver operating characteristic curve score of 0.981 and an F1 score of 0.923.

Conclusions

This study demonstrates that deep learning models can accurately distinguish between colonic adenomas with LGD and HGD. Training on a larger dataset could increase the accuracy and generalizability of the model and should be a focus of further studies.