Gastric Neoplasm Detection at Contrast-enhanced CT with Deep Learning.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Oncology and Gastroenterology
• Radiology and Diagnostic Imaging

Background:

• Gastric neoplasms pose a significant health challenge, necessitating accurate and efficient diagnostic tools.
• Current diagnostic methods for gastric neoplasms can be labor-intensive and may benefit from AI-driven enhancements.

Purpose of the Study:

• To develop and validate the Gastric Neoplasm Detection with Artificial Intelligence (GANDA) system.
• To assess GANDA's capability for automated detection, diagnosis, and segmentation of gastric neoplasms using contrast-enhanced CT scans.

Main Methods:

• A retrospective study utilizing a joint segmentation and classification 3D deep learning model (GANDA).
• Model development involved CT data from 1683 patients; validation was performed on internal, external, and real-world test cohorts.
• Performance was evaluated against board-certified radiologists using receiver operating characteristic analysis and Dice coefficient for segmentation.

Main Results:

• GANDA achieved high sensitivity and specificity for tumor detection across all test cohorts (e.g., 87.3% sensitivity, 87.2% specificity in the internal cohort).
• The model demonstrated significantly higher diagnostic accuracy than radiologists (85.3% vs. 74.2%, P = .002).
• Mean Dice coefficients for segmentation were 0.52 for gastric cancer and 0.45 for non-gastric cancer in the internal cohort.

Conclusions:

• GANDA effectively enables the detection and segmentation of gastric neoplasms in routine clinical CT scans.
• The AI approach shows promise for improving diagnostic accuracy and efficiency in gastric neoplasm screening and diagnosis.