Spatiotemporal Deep Learning on Serial CT Predicts Next-Visit Pancreatic Cancer Progression

A deep learning model using serial CT scans predicted progressive disease at the next chemotherapy follow-up visit in advanced pancreatic cancer, achieving an area under the curve of 0.77 on internal, 0.76 on external, and 0.74 on prospective test sets.

• The spatiotemporal framework integrated convolutional and LSTM neural networks to analyze 415 predicted events derived from serial CTs in 243 patients.
• Prediction performance was consistent across different chemotherapy regimens (AUC 0.68-0.79) and progression subtypes, such as target lesion growth (AUC 0.72) versus new metastases (AUC 0.77).
• Model performance was highest in patients with locally advanced disease (AUC 0.85) compared to metastatic disease (AUC 0.71).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.