Research Article: Artificial intelligence for ultrasound-based differentiation of soft tissue liposarcoma from benign masses
Abstract:
Accurate preoperative differentiation of liposarcoma from benign masses on ultrasound remains challenging and operator-dependent. We aimed to develop and validate a deep learning-based model to assist in distinguishing liposarcoma from benign masses.
In this retrospective single-center study, we collected B-mode ultrasound examinations from patients with soft tissue masses between March 2015 and March 2022 at Beijing Jishuitan Hospital affiliated to Capital Medical University. A two-stage deep convolutional neural network was developed: a U-Net model segmented tumor regions, followed by a ResNet50 classifier for tumor type prediction. The cohort ( n =?317) was split at the patient level into a development set ( n =?222) and an internal test set ( n =?95). Model performance was compared with 4 radiologists in the internal test set.
Among 317 cases, 191 were benign masses and 126 were liposarcomas. In the internal test set, the AI model achieved an AUC of 0.97 (95% CI 0.92–1.00), an accuracy of 0.97 (95% CI 0.93–1.00), a sensitivity of 0.96 (95% CI 0.86–1.00), and a specificity of 0.97 (95% CI 0.93–1.00). AI achieved significantly higher accuracy than both residents after Holm adjustment (adjusted P =?0.0156 and 0.0352, respectively). Differences between AI and either specialist were not statistically significant (both adjusted P =?1.0000).
A two-stage deep learning model using ultrasound images showed preliminary potential for differentiating liposarcoma from benign masses in this single-center cohort. External validation is required before clinical use.
Introduction:
Accurate preoperative differentiation of liposarcoma from benign masses on ultrasound remains challenging and operator-dependent. We aimed to develop and validate a deep learning-based model to assist in distinguishing liposarcoma from benign masses.
Read more