Computational pathology

Adapting tumor annotation to biopsies

Improving tumor annotation on endometrial biopsy slides through supervised fine-tuning and threshold selection.

I adapted an existing tumor-annotation model to endometrial biopsy slides using supervised fine-tuning on additional slides annotated by a pathologist. I then refined the prediction threshold to bring the automated tumor regions into closer agreement with those annotations.

Four panels compare model-predicted tumor outlines in blue with a pathologist's tumor annotations in green and other cell-type annotations in red on the same biopsy slide, at threshold settings 4, 5, 6 and 7. — Threshold comparison on one biopsy slide. Green marks the pathologist's tumor annotations, red marks another cell type, and blue outlines the model's predicted tumor regions.

Building on earlier research

Earlier work published in Nature Communications identified a p53abn-like subgroup within endometrial cancers classified as NSMP (no specific molecular profile), associated with poorer outcomes. My follow-on contribution focused on adapting the tumor-annotation stage of the image-analysis pipeline from surgical sections to biopsy slides. The model needed to handle the different tissue presentation in biopsy slides. I used additional biopsy images with pathologist-provided annotations for supervised fine-tuning, improving automated tumor annotation on these specimens.

Choosing the prediction threshold

After fine-tuning, I compared predicted tumor regions with the pathologist's annotations across threshold settings. Threshold selection further improved their agreement. The figure illustrates the trade-off: increasing the threshold removes predicted regions outside the annotated tumor, but can also exclude parts of the tumor itself.

Supervised Fine-Tuning
Domain Adaptation

Next projectRetinal disease classifier