Comparison of Convolutional Neural Networks and ChatGPT in the Evaluation of Pancreatic Tumors by EUS

Aims Endoscopic ultrasound (EUS) is the gold standard for diagnosing pancreatic tumors, and the implementation of artificial intelligence (AI) algorithms holds significant potential in this field. Different types of AI, such as Convolutional Neural Networks (CNNs) and generative AI, can be employed. Generative AI creates new, original data from learned patterns, such as text or images, while CNNs autonomously analyze and learn from large data sets. The aim of our study was to evaluate and compare the diagnostic accuracy of two AI models in characterizing pancreatic tumors through EUS.

Methods We analyzed 100 EUS images of solid pancreatic lesions to differentiate between adenocarcinoma (ADC) and pancreatic neuroendocrine tumors (PNET) using both a CNN and ChatGPT 4.0 (OpenAI, San Francisco, USA). The CNN used a ResNet architecture, trained and validated with a dataset of EUS images. ChatGPT was adapted by defining the region of interest. Histological analysis was used as the gold standard for comparing the diagnostic accuracy of each AI model.

Results The diagnostic accuracy was significantly higher for the CNN (94%) compared to ChatGPT 4.0 (88%) with a p-value<0.05. The CNN also demonstrated superior sensitivity for diagnosing ADC (98.7%), compared to 90.2% for ChatGPT (p<0.05). However, no statistically significant differences were found in the sensitivity for PNET, with values of 83.7% for the CNN and 85.7% for ChatGPT (p=0.696). Additionally, the AUC values were 0.91 for CNN and 0.88 for ChatGPT, with no significant difference between the two tests (p=0.376).

Conclusions This study demonstrates that CNNs outperform generative AI models, such as ChatGPT, in the identification of pancreatic tumors in EUS images. While ChatGPT shows limited potential in this specific context, its ability to generate synthetic data could be valuable for complementing the training of CNN models, particularly in data-scarce situations. The integration of these technologies under appropriate regulatory frameworks could contribute to the development of new diagnostic tools for pancreatic lesions using EUS, but it is essential to recognize the differences in their design and capabilities.

Publication History

Article published online:
27 March 2025