A recent study published in The Lancet Regional Health Southeast Asia journal revealed that an AI-based approach has demonstrated diagnostic performance comparable to experienced radiologists in detecting gall bladder cancer at a hospital in Chandigarh. Gallbladder cancer is known for its high aggressiveness and poor detection rates, leading to a high mortality rate. Early diagnosis is particularly challenging due to the similarities in imaging features between benign gallbladder lesions and cancerous ones, as noted by the researchers involved in the study.
To address this issue, a team from the Postgraduate Institute of Medical Education and Research (PGIMER) in Chandigarh and the Indian Institute of Technology (IIT) in New Delhi worked together to develop and validate a deep learning (DL) model for detecting gallbladder cancer using abdominal ultrasound. The goal was to assess the performance of this DL model and compare it with that of radiologists.
Deep learning is an artificial intelligence technique that enables computers to process data in a manner inspired by the human brain. In a recent study conducted at PGIMER, a tertiary care hospital, abdominal ultrasound data from patients with gallbladder lesions was utilized.
The study involved training a deep learning (DL) model on a dataset consisting of 233 patients. Additionally, the model was validated on a separate group of 59 patients and subsequently tested on another set of 273 patients.
To assess the DL model's performance, various metrics such as sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC) were used. The AUC is a widely utilized measure for evaluating the accuracy of diagnostic tests.
In addition to the DL model, two radiologists independently reviewed the ultrasound images, allowing a comparison of their diagnostic performance to that of the DL model.
According to the study, the DL model performed well in the test set for detecting gallbladder cancer (GBC). It had a sensitivity of 92.3%, which means it correctly identified 92.3% of the positive cases. The specificity of the model was 74.4%, indicating that it correctly identified 74.4% of the negative cases. The AUC (Area Under the Curve) value, which is a measure of the model's overall performance, was 0.887. These results were comparable to the performance of radiologists in detecting GBC.
The researchers found that the DL-based approach exhibited high sensitivity and AUC for detecting GBC even in challenging scenarios. It performed well in the presence of stones, contracted gallbladders, small lesion sizes (less than 10 mm), and neck lesions. The performance of the DL model in these cases was also comparable to that of radiologists.
In the test set, the DL model had a sensitivity of 92.3 per cent, specificity of 74.4 per cent, and an AUC of 0.887 for detecting GBC, which was comparable to both radiologists, according to the study.
The DL-based approach showed high sensitivity and AUC for detecting GBC in the presence of stones, contracted gallbladders, small lesion size (less than 10 mm), and neck lesions, which were also comparable to the radiologists, the researchers said.
The DL model exhibited higher sensitivity for detecting the mural thickening type of GBC compared to one of the radiologists, despite a reduced specificity, they said.
"The DL-based approach demonstrated diagnostic performance comparable to experienced radiologists in detecting GBC using ultrasound," the authors of the study noted.
"Further multicentre studies are recommended to fully explore the potential of DL-based GBC diagnosis," they added.
(With inputs from PTI)