Endoscopic ultrasound-guided fine-needle biopsy (EUS-FNB) has become the standard method for obtaining tissue from pancreatic tumors before surgery. However, the tissue samples obtained this way are small, fragmented, and mixed with blood and inflammatory cells, making accurate pathological diagnosis extremely challenging.
Isolated cancer cell clusters, called invasive ductal carcinoma components (ICCs), are the hardest to identify because they appear as small groups of cells without the typical glandular architecture. Even experienced pancreatic pathologists can miss these cells in difficult specimens.
The research team at Kurume University collected 594 whole-slide images (WSIs) from EUS-FNB specimens of pancreatic cancer patients. Expert pancreatic pathologists carefully annotated all cancer cells in a training set of 412 slides, marking both the glandular invasive duct carcinoma (IDC) and the isolated invasive carcinoma components (ICC).
The deep learning model used EfficientNet-B1, a convolutional neural network architecture that balances performance with computational efficiency. The model was trained using a combination of transfer learning from ImageNet and fully supervised learning on the annotated biopsy slides.
The model was evaluated on a test set of 120 WSIs whose diagnoses were agreed upon by three independent pancreatic pathologists. Cases where pathologists disagreed were excluded from evaluation to ensure the ground truth was reliable.
The model achieved an AUC of 0.984 on the test set, with accuracy of 94.2%, sensitivity of 93.0%, and specificity of 97.1%. These results are comparable to or better than previously published AI models for other cancer types on whole-slide images.
Notably, the model was able to accurately detect isolated cancer cell clusters (ICCs) among abundant blood and fibrin background, which represents the most diagnostically difficult scenario for human pathologists. This is the scenario where AI assistance could have the greatest clinical impact.
False negatives tended to occur in cases where the area of cancer cell foci was very small. False positives were rare and mainly involved contaminating gastric gland tissue that passed through the stomach during the biopsy needle path.
The proposed use case is not to replace pathologists but to provide a diagnostic support system for difficult EUS-FNB specimens. When a general pathologist encounters a specimen with only isolated cancer cells and abundant background contamination, an AI second opinion could confirm or prompt a more careful look.
This is particularly valuable at hospitals without dedicated pancreatic pathologists. Most biopsies are performed at academic centers, but community hospitals increasingly perform EUS-FNB and may not have the subspecialty expertise to reliably interpret the most difficult specimens.
To the authors' knowledge, this is the first study to develop and validate a deep learning model specifically for detecting adenocarcinoma on EUS-FNB specimens. Prior work focused on surgical resection specimens, which are much larger and easier to analyze.
Future work should train and validate the model across multiple institutions to assess generalizability. The current model was trained on specimens from a single hospital, and performance on specimens from different scanners or biopsy techniques may vary. Expanding the training dataset with more ICC cases would also likely improve false-negative rates for the smallest cancer foci.