Vision Transformer Model Shows Promise for Pancreatic Tumor Segmentation in Endoscopic Ultrasound Images

Global: Vision Transformer Model Shows Promise for Pancreatic Tumor Segmentation in Endoscopic Ultrasound Images

A new study published on arXiv evaluated a deep‑learning segmentation model that leverages a Vision Transformer architecture to identify pancreatic tumors in endoscopic ultrasound (EUS) images. Researchers trained the model on 17,367 images drawn from two publicly available datasets and assessed its performance through five‑fold cross‑validation and an independent external test set.

Model Architecture and Training

The segmentation system was built within the USFM framework, employing a Vision Transformer as the backbone. Pre‑processing steps included conversion to grayscale, cropping, and resizing each image to 512 × 512 pixels before feeding them into the network.

Cross‑Validation Performance

During five‑fold cross‑validation, the model achieved a mean Dice similarity coefficient (DSC) of 0.651 ± 0.738 and an intersection‑over‑union (IoU) of 0.579 ± 0.658. Sensitivity was recorded at 69.8%, specificity at 98.8%, and overall accuracy at 97.5%.

External Validation Results

When tested on an independent set of 350 EUS images manually segmented by radiologists, the model produced a DSC of 0.657 (95% CI: 0.634‑0.769) and an IoU of 0.614 (95% CI: 0.590‑0.689). Sensitivity rose to 71.8% while specificity remained high at 97.7%.

Observed Errors

Analysis indicated that 9.7% of cases contained erroneous multiple predictions, highlighting specific scenarios where the algorithm struggled to delineate tumor boundaries.

Study Limitations

The authors noted that heterogeneity across the source datasets and the relatively small size of the external validation cohort may limit the generalizability of the findings.

Implications and Future Work

While the results suggest that Vision Transformer‑based segmentation can support radiologists in EUS‑guided diagnosis, the study calls for further refinement, larger multi‑center validations, and prospective clinical trials to confirm utility in routine practice.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.