Breast-to-Uterine Tumor Transfer Learning and Data Augmentation: Addressing Limited Ultrasound Data for Uterine Malignancy Classification

Uterine cancer is among the most common cancers suffered by women worldwide. Delayed diagnosis of its malignancy diminishes the rate of successful treatment outcomes. To mitigate this risk, deep learning has demonstrated its considerable potential in assisting clinicians with medical image analysis. Nevertheless, the development of reliable deep learning models for uterine tumor diagnosis is hindered by the limited availability of uterine ultrasound datasets, as such models typically require large amounts of data. In previous studies, to overcome this limitation, researchers often use data augmentation or transfer learning. However, cross-organ transfer learning-specifically from breast to uterine ultrasound images-remains largely underexplored. This study explores a comparative analysis of three approaches-data augmentation, single-stage transfer learning (ImageNet), and multi-stage transfer learning (ImageNet to breast to uterus)-to overcome limited data availability in uterine ultrasound images when it is applied to EfficientNet B0. Multi-stage transfer learning is further evaluated via fine-tuning and feature extraction. Results indicate that multi-stage transfer learning with feature extraction has the best accuracy (85%), outranking singlestage ImageNet transfer learning (77%), and data augmentation alone is less than 50% accurate. Importantly, the use of data augmentation consistently degrades transfer learning performance, probably because of the distortion of clinically relevant image features, leading to misdiagnosis and less reliable results in clinical applications. Finally, we find that multi-stage transfer learning with feature extraction is the best method for the diagnosis of uterine tumor malignancy under limited data conditions.