← Back to Home

Distinguishing Precancerous Colorectal Lesions: Vision Transformer Improves Detection Sensitivity in Pathology Image Testing

A study using a public dataset found that a Vision Transformer achieved nearly 70% sensitivity in identifying sessile serrated adenomas, compared with about 40% for the convolutional model used for comparison. Whether this advantage can translate into support for clinical interpretation still requires validation across datasets and in prospective studies.

By SURL BioNews

The risk associated with colorectal polyps sometimes lies in subtle differences in tissue structure. Sessile serrated adenomas (SSAs), which are precancerous lesions, can resemble hyperplastic polyps on pathology slides. A study recently highlighted by Northeastern University in the United States suggests that switching to a Vision Transformer architecture may help artificial intelligence models identify these easily confused lesions more effectively.

The research comes from the Amal Precision Medicine Laboratory at Northeastern University’s Roux Institute. The findings were first published online on August 12, 2026, in the conference proceedings series *Lecture Notes in Computer Science*; a university report in October and its republication by Medical Xpress brought renewed attention to the study. It evaluated the classification of existing histopathology images and was not a trial of real-time lesion detection during colonoscopy.

The team used 3,152 hematoxylin and eosin (H&E)-stained images from the public MHIST dataset to compare three Transformers—ViT, CaiT, and DeiT—with three convolutional neural networks (CNNs): ResNet-18, 34, and 50. All models underwent the same stratified five-fold cross-validation, taking turns training and being evaluated within the same dataset to provide a common basis for comparing the architectures.

According to the paper’s abstract, ViT achieved 69.7% sensitivity in identifying SSAs, compared with 39.2% for the best-performing CNN in the comparison—a difference of 30.5 percentage points. In other words, ViT identified more images that were actually SSAs in this image test. Its area under the ROC curve was 0.917, compared with 0.887 for the CNN; the university also reported overall accuracies of approximately 86% and 80%, respectively. These figures reflect differences between the models and do not demonstrate that AI outperforms pathologists.

Sensitivity is particularly important here because approximately 71% of the images in the dataset depict hyperplastic polyps. When one image category accounts for the majority, a model biased toward predicting that category may achieve seemingly reasonable overall accuracy while missing less common precancerous lesions. The study found that Transformers maintained more balanced classification performance even without deliberately increasing how often the minority class was sampled. Study lead Saeed Amal suggested that this may be related to the models’ ability to integrate relationships between different regions of an image.

However, sensitivity approaching 70% also means that approximately 30% of SSA images were still not correctly identified in this test. More critically, five-fold cross-validation was still conducted within a single dataset and cannot replace external validation. Whether the models can maintain their performance when faced with differences in staining, scanning, and cases at other hospitals remains unanswered; the study also did not establish that they can reduce patients’ risk of developing cancer later.

According to the university report, the team is working with pathologists and gastroenterologists toward prospective studies to assess the models’ accuracy and usefulness in actual clinical workflows. The next priority is to determine whether this advantage in image classification can help physicians miss fewer lesions, while clarifying how much additional review work misclassifications would create, to assess how far the technology remains from becoming a reliable pathology support tool.

References

  1. Medical Xpress / Northeastern University
  2. Springer Nature / Lecture Notes in Computer Science
  3. Northeastern Global News