Advancing discriminative to representative: A self-supervised classification framework by integrating selective diffusion model learned features for uterine adenomyosis diagnosis in transvaginal ultrasound image

Deep learning based methods have attracted attention for uterine adenomyosis in transvaginal ultrasound (TVUS) images due to their powerful abilities of feature representation, yet are hindered by a lack of labeled data. Self-supervised learning methods show the potential for this problem since they can learn feature representation from abundant unlabeled data for downstream tasks. However, the differences in TVUS images are usually subtle and ambiguous, while existing self-supervised learning methods tend to disrupt discriminative information when using auxiliary tasks such as generating positive–negative image pairs and reconstructing masked images. To address these problems, in this paper, we propose a self-supervised classification framework that derives the representative feature from diffusion model learned features and complements it with detailed texture information of the image. Specifically in the pre-training stage, the U-Net is trained following the Denoising Diffusion Probabilistic Model (DDPM) so that most of the valuable discriminative information is preserved by the network features. In the classification stage, the cross-attention fusion (CAF) module is designed to integrate multiple selective features from the pre-trained U-Net, yielding a representative feature that contains more comprehensive discriminative information. Furthermore, the adaptive normalization (AN) module is designed to introduce detailed texture information from parallel convolutional neural network (CNN) into the representative feature, enhancing the ability to recognize subtle differences between TVUS images. The proposed method is evaluated on the self-collected TVUS dataset and compared with state-of-the-art deep learning models. Experimental results demonstrate that the proposed method significantly enhances the accuracy, specificity, and sensitivity of uterine adenomyosis diagnosis, indicating that the network is capable of achieving accurate and robust diagnosis of uterine adenomyosis with a limited labeled dataset, thereby advancing the clinical application of deep learning for adenomyosis diagnosis.