ABSTRACT
Background: Bone scans are often used to identify bone metastases, but their low specificity may necessitate further studies. Deep learning models may improve diagnostic accuracy but require both medical and programming expertise. Therefore, we investigated the feasibility of constructing a deep learning model employing ChatGPT for the diagnosis of bone metastasis in bone scans and to evaluate its diagnostic performance. Method: We examined 4626 consecutive cancer patients (age, 65.1 ± 11.3 years; 2334 female) who had bone scans for metastasis assessment. A nuclear medicine physician developed a deep learning model using ChatGPT 3.5 (OpenAI). We employed ResNet50 as the backbone network and compared the diagnostic performance of four strategies (original training set, original training set with 1:10 class weight, 10-fold data augmentation for positive images only, and 10-fold data augmentation for all images) to address the class imbalance. We used a class activation map algorithm for visualization. Results: Among the four strategies, the deep learning model with 10-fold data augmentation for positive cases only, using a batch size of 16 and an epoch size of 150, achieved the area under curve of 0.8156, the sensitivity of 56.0 %, and specificity of 88.7 %. The class activation map indicated that the model focused on disseminated bone metastases within the spine but might confuse them with benign spinal lesions or intense urinary activity. Conclusions: Our study illustrates that a clinical physician with rudimentary programming skills can develop a deep learning model for medical image analysis, such as diagnosing bone metastasis in bone scans using ChatGPT. Model visualization may offer guidance in enhancing deep learning model development, including preprocessing, and potentially support clinical decision-making processes.
