A new AI pipeline just got meaningfully better at telling colorectal cancer tumor grades apart in biopsy slides.
Researchers built an automated segmentation system for whole-slide histopathology images that labels tissue by tumor grade (1 through 3) or normal mucosa. It runs transformer-based image models over overlapping patches of the slide, applies test-time augmentation, and uses a large language model to help tune how training data gets augmented along the way. The best-performing models were then combined through soft voting - essentially averaging their predictions - and cleaned up with standard image post-processing like Gaussian blurring and removal of small stray detections. On a colorectal cancer grading dataset, the approach raised the F1 score, a measure that balances false positives and false negatives, from 62.92 to 69.84.
Colorectal cancer is the third most common cancer and the deadliest of the gastrointestinal cancers, and diagnosis still hinges on a pathologist manually grading biopsy tissue under a microscope - slow, and subject to human variation. Tools like this aim to pre-flag likely tumor regions and grades before a pathologist ever looks at the slide. The more interesting wrinkle is the method itself: using one AI model to help design the training process for another is a trend showing up well beyond text generation, in a field as unglamorous as histopathology.
A roughly seven-point F1 gain on one benchmark dataset is a real result, not a breakthrough. It is a lab number on a research dataset, not something that has survived clinical trials, regulatory review, or contact with a messy real-world pathology lab - the usual places these systems go to die before reaching a hospital.