Artificial Intelligence (AI)-Assisted Systems for Pneumothorax Detection: Latest Cureus Update

This proof-of-concept study was published in December 2025 in the Journal Cureus

Introduction

Artificial intelligence (AI) systems for detecting pneumothorax are being improved. They can be faster and more accurate than doctors reading radiographs. However, most AI tools only work with images. For instance, a neural network model performed well on test images but not on real hospital images. This highlights the challenge of making AI work across different settings and avoiding such errors. New AI tools on platforms such as Google Cloud Vertex AI work well, even for small pneumothorax cases. They can provide a second opinion. In addition to images, new AI models incorporate text, other data, and even audio. These models may assist with report writing and answering radiology questions. A survey showed that while these models exhibit potential, they also encounter challenges, including inadequate data, inaccuracies, and usability issues.

Study Overview

The authors conducted a preliminary comparative evaluation of two general-purpose multimodal models, GPT-4o and Gemini 2.5 Pro, for detecting pneumothorax on chest radiographs. This study aimed to benchmark the strengths and limitations of these models and assess the potential clinical contexts in which each approach may be valuable. The evaluation involved 2,000 frontal chest radiographs, evenly divided between cases with and without clinically diagnosed pneumothorax.

Both models were queried following a standardized prompt: "Given a frontal chest radiograph, analyze and determine evidence of pneumothorax. Look for visible pleural line without lung markings, deep sulcus sign, lung translucency asymmetry, and collapse signs.

• The models generated binary predictions for each radiograph: 0 (no pneumothorax) or 1 (pneumothorax present).
• The predictions were compared with the ground-truth labels to generate confusion matrices showing true positives, true negatives, false positives, and false negatives.
• The performance metrics included accuracy, precision, recall, and F1 score.
• Confidence intervals (95%) were used for accuracy and recall, and bootstrapping was used for F1 scores.

Key findings:

Table 1 shows the comparative performance of the GPT-4o and Gemini models for pneumothorax detection. The matrix highlights GPT-4o’s more balanced performance, with moderate precision and specificity but a higher number of false negatives, which raises concerns for clinical triage scenarios where missed cases are critical. The matrix demonstrates Gemini’s substantially higher recall, indicating stronger sensitivity and reduced precision compared with GPT-4o, reflecting its emphasis on minimizing missed cases of pneumothorax. However, fewer false negatives make it potentially more suitable for early screening contexts, although this comes at the cost of increased false positives and lower precision.

Table 1: Comparative performance of GPT-4o and Gemini model for pneumothorax detection (2,000 chest X-rays - 1,000 pneumothorax and 1,000 nonpneumothorax)

Performance Metrics	GPT-4o model	Gemini model
Overall Accuracy	64%	62%
Precision	66%	55%
Recall	57%	88%
F1 Score	61%	68%
True Positives	571 Pneumothorax cases	879 Pneumothorax cases
True Negatives	710 no-pneumothorax cases	275 no-pneumothorax cases
False Positives	290 Pneumothorax cases	121 Pneumothorax cases
False Negatives	429 no-pneumothorax cases	725 no-pneumothorax cases

Potential Clinical Implications

The authors highlighted the complementary strengths of GPT-4o and Gemini 2.5 Pro in pneumothorax detection. Pneumothorax is an urgent condition in which delayed recognition can lead to cardiorespiratory collapse. Early identification on chest radiographs, especially in emergency settings, is vital. AI assistance may expedite the detection and reduce errors during high-workload periods. GPT-4o's balanced performance seems suitable for general screening, whereas Gemini's high sensitivity is suitable for triage workflows, where minimizing missed diagnoses is critical. This shows the need to tailor AI deployment to the clinical context. Regulatory preparedness, validation, integration strategies, and human supervision are essential before clinical deployment.

Reference: Chetla N, Patel S, Sharma S, et al. Evaluating AI Models for Pneumothorax Detection on Chest Radiographs: Diagnostic Accuracy and Clinical Trade-Offs. Cureus 17(12): e99298. Published 2025 Dec 15. DOI 10.7759/cureus.99298