ChatGPT-assisted deep learning for diagnosing bone metastasis in bone scans: Bridging the AI Gap for Clinicians

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.

Keywords: Bone metastasis; Bone scan; ChatGPT; Convolutional neural network; Deep learning.

Graphical abstract

Fig. 1

Flow diagram of patient enrollment.

Fig. 2

Dataset composition. Five different 8:2 training sets and test sets were constructed not to overlap the test sets.

Fig. 3

Overview of the deep learning model to classify bone metastasis positive and negative of bone scans. After unifying the intensities of the bone scan images in the preprocessing stage, the diagnostic performance in the test set is compared after training with four different strategies.

Fig. 4

Representative cases of the final model. True-positive cases show disseminated bone metastases weighted in the spine, especially the T-spine (a, b). However, patients with a single lesion in the T-spine (c, white arrowhead) and bone metastases outside the spine (d, white arrowhead) are classified as false negatives. There is no significant region of interest in the trained model in the entire skeleton in true-negative cases (e, f). However, a traumatic compression fracture (g, white arrowhead) and large intense urinary bladder activity (h, white arrowhead) can be classified as false positives. The region outside the body can be the area of interest for the trained model (d, f, g, h).