Yuzhong Zhao (赵煜中)

The diverse tasks of X-ray bone image analysis necessitate training and maintaining numerous task-specific models. Recent advancements in foundation models have attracted extensive attention across various fields, promising significant improvements in medical applications. These models depend on large-scale datasets and utilize the self-supervised learning framework for their training. However, the study for the specific scenario when data acquisition is challenging still remains limited, such as the modality of bone X-ray images. This limitation is partly due to the limited availability of large-scale and curated bone X-ray datasets for research and model development. In this paper, we aim to fill this gap by applying DINO and DINOv2 frameworks for self-supervised learning across two bone X-ray image datasets. Our comprehensive analysis covers a range of downstream tasks, including supervised regression for accurate bone age assessment, unsupervised clustering through a Gaussian mixture model, and supervised regions of interest (ROI) localization. Our experiments demonstrate that pretrained models can significantly improve the downstream task performance and the extracted features conform to the property of bone age. The study shows the effectiveness of self-supervised learning on the bone X-ray images, providing insights into their capabilities to improve diagnostic precision and generalizability in medical imaging.

The rapid advancement of deep learning in medical imaging necessitates reliable out-of-distribution (OOD) detection to ensure safe clinical deployment and prevent potential patient harm. Traditional supervised methods, which rely on closed-world assumptions, struggle in real-world scenarios where abnormal samples are both diverse and scarce. To address this, we propose a novel synthetic outlier-based OOD detection framework that distinguishes OOD samples without requiring real abnormal data. Our method generates a diverse and structurally plausible set of pseudo-outliers through a hybrid synthesis pipeline combining local and global transformations. By incorporating these synthetic outliers into a constrained optimization framework during training, our model learns a more discriminative decision boundary between in-distribution (ID) and OOD samples. We conducted extensive evaluations on four diverse medical imaging datasets, and the results show that our approach significantly outperforms state-of-the-art techniques in medical OOD detection while preserving high in-distribution classification accuracy. Our findings demonstrate that strategically generating synthetic outliers provides an effective and practical solution for enhancing model robustness in safety-critical medical applications.

Bone age assessment (BAA) is a widely used clinical practice for the biological development of adolescents. The Tanner Whitehouse (TW) method is a traditionally mainstream method that manually extracts multiple regions of interest (ROIs) related to skeletal maturity to infer bone age. In this paper, we propose a deep learning-based method for fully automatic ROIs localization and BAA. The method consists of two parts: a U-net-based backbone, selected for its strong performance in semantic segmentation, which enables precise and efficient localization without the need for complex pre- or post-processing. This method achieves a localization precision of 99.1% on the public RSNA dataset. Second, an InceptionResNetV2 network is utilized for feature extraction from both the ROIs and the whole image, as it effectively captures both local and global features, making it well-suited for bone age prediction. The BAA neural network combines the advantages of both ROIs-based methods (TW3 method) and global feature-based methods (GP method), providing high interpretability and accuracy. Numerical experiments demonstrate that the method achieves a mean absolute error (MAE) of 0.38 years for males and 0.45 years for females on the public RSNA dataset, and 0.41 years for males and 0.44 years for females on an in-house dataset, validating the accuracy of both localization and prediction.

Anomaly detection from medical images is an important task for clinical screening and diagnosis. In general, a large dataset of normal images are available while only few abnormal images can be collected in clinical practice. By mimicking the diagnosis process of radiologists, we attempt to tackle this problem by learning a tractable distribution of normal images and identify anomalies by differentiating the original image and the reconstructed normal image. More specifically, we propose a normalizing flow-based autoencoder for an efficient and tractable representation of normal medical images. The anomaly score consists of the likelihood originated from the normalizing flow and the reconstruction error of the autoencoder, which allows to identify the abnormality and provide an interpretability at both image and pixel levels. Experimental evaluation on two medical images datasets showed that the proposed model outperformed the other approaches by a large margin, which validated the effectiveness and robustness of the proposed method.

Anomaly detection of Positron Emission Tomography (PET) are important tasks for clinical diagnosis and treatments. For 3D PET images, it is arduous for the annotations of lesions and the number of positive samples are generally very limited. Although there have been many work on deep learning based anomaly detection for medical images, most of them cannot be directly applied on 3D PET data. In this paper, from a set of normal PET data, we propose a tractable and efficient 3D normalizing flow based model, namely PET-3DFlow for anomaly detection and provide lesion localization for the reference of clinicians. The loss of In PET-3DFlow consists of the log-likelihood originated from the normalizing flow model and the reconstruction error of the autoencoder learned from the normal images. Experimental evaluation on 3D PET data showed that the proposed model outperformed the other approaches, which validated the effectiveness of our proposed PET anomaly detection method.