Automatic feature extraction by supervised and contrastive self-supervised learning based on wavelet and hard negatives to detect HIFU lesion area

Abstract

The adoption of Deep Neural Networks has surged due to their ability to automatically extract features and employ diverse approaches in data analysis. This research proposes a novel feature extraction method that doesn't rely on labeled training data, particularly considering the utilization of hard negatives. Given the remarkable success of DNN-based models in analyzing various medical images, including disease diagnosis and detection, this paper delves into diagnosing the lesion area against the normal area, particularly in the context of the non-invasive treatment of HIFU. Monitoring and analyzing inputs related to the lesion area are crucial to prevent damage to normal tissue during the heating process. However, several challenges exist in ultrasound medical imaging, including small sample sizes, data lacking labels, and the time-intensive nature of deep supervised training. These challenges have motivated the introduction of a new self-supervised deep learning method. While supervised learning excels in accuracy, unlabeled data holds valuable information discarded in supervised approaches. Conversely, ultrasonic data's nature lies in the RF signal, offering a detailed acoustic structure of tissue. Acknowledging the limitations and advantages of each method, an effective approach leveraging both signal and image simultaneously is presented. This integrated method enhances diagnostic capabilities and contributes to improve monitoring of HIFU procedures. The proposed methodology for classifying HIFU lesion areas attained high performance metrics: 95% accuracy, 94% precision, 96% recall, and a 95% F1-score. These outcomes underscore the efficacy of the proposed method in accurately classifying HIFU lesion areas.