Benchmarking CNN architectures using transfer learning for Lipomatous tumor classification: a comprehensive study

Abstract

This study evaluates five pre-trained CNN architectures—ResNet50, DenseNet121, EfficientNetB3, InceptionV3, and MobileNetV2—for distinguishing benign lipoma from well-differentiated liposarcoma (WDLPS) using T1-weighted MRI. Using a uniform frozen transfer-learning framework and MRI data from 115 patients in the WORC database, MobileNetV2 achieved the best overall performance, with 96.7% accuracy and 0.997 AUROC. DenseNet121 showed strong WDLPS recall and may be preferable for safety-critical screening. Although EfficientNetB3 achieved the highest WDLPS recall, its high false-positive rate and poor overall performance made it unsuitable under the tested configuration. The findings support MobileNetV2 as an effective, lightweight architecture for automated MRI-based lipoma and WDLPS classification.