Google AI Studio shows potential for detecting lung cancer on CT imaging, but its “interpretative oversensitivity and limited specificity” underscore the need for human oversight, according to research shared at the recent ECR in Vienna.
Presenter Zainab Dawood Aljneibi of Fatima College of Health Sciences in Abu Dhabi, United Arab Emirates, and colleagues found that the AI model could identify lung malignancy and produce consistent, structured reports, but that it needs “further refinement to reduce misclassification in benign and normal presentations.”
The group conducted a study in March 2025 that included information taken from the publicly available IQ-OTH/NCCD dataset. The dataset consisted of 110 CT cases, of which 55 were normal, 15 were benign, and 40 were malignant. The investigators refined a pretrained convolutional neural network in Google AI Studio using a subset of images. They tracked quantitative measures such as diagnostic accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), as well as qualitative measures such as report structure, terminology use, and error patterns.
Overall, Aljneibi and colleagues reported that Google AI Studio achieved an overall accuracy of 75.5% for identifying lung cancer, with a sensitivity of 74.5% and a specificity of 76.4%. The model missed 14 positive cases and produced 13 false positives.
They also found the following:
- Google AI Studio identified malignant cases with “high discriminative performance” (AUC = 0.9), but benign cases were more difficult for it to classify (AUC = 0.62).
- Qualitative analysis showed that Google AI Studio consistently used radiological terminology, with structured descriptions that resembled human reporting.
- In normal cases, Google AI Studio’s use of terms such as “clear lung fields” confirmed its accurate recognition of normal results, but occasional references to ground-glass opacities reflected “oversensitivity to minor changes,” the group noted.
- Google AI Studio described benign cases with bilateral opacities and fibrotic changes, but “overlapping terminology sometimes led to misclassification.”
- Google AI Studio described malignant cases with features such as “mass” and “lesion” — highlighting “effective feature recognition but limited specificity in non-malignant contexts,” the investigators explained.
Aljneibi noted that the research identified four areas of improvement for use of Google AI Studio: addressing overreliance on nonspecific features, diversifying training data to improve generalizability, integrating radiomics to balance sensitivity and specificity, and “promoting continuous human-AI collaboration.”
“Human oversight [of Google AI Studio] is essential for clinical use,” she concluded.
Full coverage of ECR 2026 can be found here.