AI Gastric Cancer Detection | CT Scan Screening
- A novel artificial intelligence (AI) model called GRAPE shows promise in detecting gastric cancer (GC), or stomach cancer, through noncontrast computed tomography (CT) scans.
- The study, approved by a centralized institutional review board (IRB-2024-279, Clinicaltrials.gov registration NCT06614179), involved three autonomous cohorts.
- Inclusion criteria for the GC group required a confirmed diagnosis via gastroscopic pathological biopsy and a pre-treatment noncontrast CT scan.
Harnessing the power of artificial intelligence, a groundbreaking model called GRAPE is showing remarkable promise in early gastric cancer (GC) detection.This innovative AI analyzes noncontrast CT scans, offering a potential game-changer for stomach cancer screening and diagnosis. The system aims to improve early diagnosis and treatment planning and utilizes a two-stage strategy, localizing the stomach region and then performing tumor segmentation. With performance evaluated using metrics like AUC and specificity, this advancement is poised to reshape how we approach the primarykeyword. This study involved rigorous testing across diverse patient cohorts. This work is brought to you by news directory 3. How will this innovative secondarykeyword change cancer care? Discover whatS next …
AI Spots Stomach Cancer Early on CT Scans
A novel artificial intelligence (AI) model called GRAPE shows promise in detecting gastric cancer (GC), or stomach cancer, through noncontrast computed tomography (CT) scans. The AI system aims to improve early diagnosis and treatment planning.
The study, approved by a centralized institutional review board (IRB-2024-279, Clinicaltrials.gov registration NCT06614179), involved three autonomous cohorts. A training cohort included 3,470 GC patients and 3,250 non-GC (NGC) participants,aged 18-99,enrolled from September 2006 to June 2024 across two centers in China. An internal validation cohort, also in China, consisted of 650 GC and 648 NGC participants, aged 20-94, enrolled between December 2006 and April 2024.The external validation cohort, conducted from January 2011 to August 2024, included 18,160 participants aged 18-80 from 16 centers who underwent gastroscopic examination.
Inclusion criteria for the GC group required a confirmed diagnosis via gastroscopic pathological biopsy and a pre-treatment noncontrast CT scan. NGC participants were confirmed as not having GC by gastroscopy and had a noncontrast CT examination within six months of the gastroscopy, or were confirmed without GC based on a one-year follow-up after a noncontrast CT scan. Researchers reviewed medical records, noting age, sex, T stage and location.
GRAPES performance was also assessed in real-world hospital opportunistic screening using two cohorts: 41,178 participants with noncontrast CT scans (2018-2024) from two regional hospitals, and 37,415 participants with recent noncontrast CT scans (2022-2024) from a cancer center. In the first hospital,about 45% of patients were from outpatient,emergency,or physical examination departments,while the rest were inpatients. The second hospital saw a similar split.The cancer center excluded 4,573 patients with pre-existing GC diagnoses; roughly 31% were outpatients, and 39% were inpatients.
CT images were collected before gastroscopy. Stomach segmentation was achieved using a semi-supervised self-training approach, combining a publicly available dataset with manually annotated masks and an internal training set. Tumor segmentation was performed by two radiologists using ITK-SNAP software. Discrepancies were resolved through discussion. The DEEDS registration algorithm aligned venous-phase images with noncontrast images to obtain tumor annotations on noncontrast images, which were then delineated and confirmed.
The GRAPE model analyzes 3D noncontrast CT scans to detect and segment GC, producing a pixel-level segmentation mask and a classification score. It uses a two-stage strategy: first,localizing the stomach region using a segmentation network (nnUNet),then cropping the region for tumor segmentation and patient classification. The model integrates subtle 3D morphometric patterns and contextual radiological features to detect early GCs, even with the resolution limitations of noncontrast CT.
To handle variations in CT imaging parameters, scans were resampled to a uniform voxel spacing and intensity values were normalized. During training, fixed-size 3D patches were extracted, and Gaussian noise was added. At test time,a sliding window strategy ensured consistent performance.
A fivefold cross-validation approach was used, with ensemble learning averaging classification probabilities. GRAPE addresses stomach filling variability through training on a range of gastric volumes. The model’s interpretability is enhanced by segmentation outputs and grad-CAM visualization of convolutional feature maps.
The primary goal was binary classification of GC, with a GRAPE score above 0.5 indicating high risk. The study assessed GC detection by T stage, TNM stage, lesion location and stomach filling during CT examination.
A reader study compared GRAPE’s performance with 13 radiologists (5 senior, 8 junior). In one session, GRAPE was compared to radiologists; in another, radiologists were provided with GRAPE’s predictions. A one-month washout period separated the sessions.
Performance was evaluated using AUC,sensitivity,specificity,positive predictive value and balanced accuracy,with confidence intervals calculated via bootstrap replications. Permutation tests persistent the significance of sensitivity,specificity and balanced accuracy comparisons (P < 0.05). Data analysis was conducted in Python using numpy, scipy and scikit-learn packages.