Deformable image registration is crucial for medical image analysis, yet the complexity of deep learning networks often limits their deployment on resource-limited devices. Current distillation methods in registration tasks fail to effectively transfer complex deformation handling capabilities to non-iterative lightweight networks, leading to insignificant performance improvement. To address this, we propose the Framework for Iterative to Non-iterative Distillation (FIND), which efficiently transfers these capabilities to a Non-Iterative Lightweight (NIL) network. FIND employs a dual-step process: first, using recurrent distillation to derive a high-performance non-iterative teacher assistant from an iterative network; second, using advanced feature distillation from the assistant to the lightweight network. This enables NIL to perform rapid, effective registration on resource-limited devices. Experiments across four datasets show that NIL can achieve up to 60 times faster performance on CPU and 89 times on GPU than compared deep learning methods, with superior registration accuracy improvements of up to 3.5 points in Dice scores.

Image registration of liver dynamic contrast-enhanced computed tomography (DCE-CT) is crucial for diagnosis and image-guided surgical planning of liver cancer. However, intensity variations due to the flow of contrast agents combined with complex spatial motion induced by respiration brings great challenge to existing intensity-based registration methods. To address these problems, we propose a novel structure-aware registration method by incorporating structural information of related organs with segmentation-guided deep registration network. Existing segmentation-guided registration methods only focus on volumetric registration inside the paired organ segmentations, ignoring the inherent attributes of their anatomical structures. In addition, such paired organ segmentations are not always available in DCE-CT images due to the flow of contrast agents. Different from existing segmentation-guided registration methods, our proposed method extracts structural information in hierarchical geometric perspectives of line and surface. Then, according to the extracted structural information, structure-aware constraints are constructed and imposed on the forward and backward deformation field simultaneously. In this way, all available organ segmentations, including unpaired ones, can be fully utilized to avoid the side effect of contrast agent and preserve the topology of organs during registration. Extensive experiments on an in-house liver DCE-CT dataset and a public LiTS dataset show that our proposed method can achieve higher registration accuracy and preserve anatomical structure more effectively than state-of-the-art methods.

Achieving the pathological type diagnosis of pulmonary nodules on chest CT is a critical step in the early detection of lung cancer and treatment of patients. Based on a small and unbalanced self-constructed dataset, we achieved intelligent diagnosis of five pathological types including adenocarcinoma, squamous cell carcinoma, small cell carcinoma, inflammatory and other benign diseases for the first time. In order to reduce the dependence of deep convolutional neural network (DCNN) on a large amount of training data, a reverse adversarial classification network (RACN) was proposed based on semi-supervised learning, which consists of a reverse generative adversarial network (RGAN) for unsupervised regression and a supervised classification network (CN). In RGAN, five specific normal distributions P with different means and variances were assigned to represent the five pathological types, and then a special regression task was designed by mapping pulmonary nodules to the random sampling Z of P. The input of generator in RGAN is set to 3D nodule volume data, the inputs of discriminator are set to Z and the output of generator. The regression task enables RGAN to extract specific features, which will be deeply integrate into CN to improve the classification performance. Experiments showed that the average sensitivity of RACN in detecting malignant nodules was 0.6525, where the sensitivity of adenocarcinoma, small cell carcinoma and squamous cell carcinoma was 0.8426, 0.5604 and 0.5543. Besides, the RACN can achieve 93.21% accuracy for diagnosing malignant nodules on the public LIDC-IDRI dataset, obtaining the state-of-the-art results.

To accurately diagnose pneumonia patients on a limited annotated chest X-ray image dataset, a prior knowledge-based active attention network (PKA2-Net1) was constructed. The PKA2-Net uses improved ResNet as the backbone network and consists of residual blocks, novel subject enhancement and background suppression (SEBS) blocks and candidate template generators, where template generators are designed to generate candidate templates for characterizing the importance of different spatial locations in feature maps. The core of PKA2-Net is SEBS block, which is proposed based on the prior knowledge that highlighting distinctive features and suppressing irrelevant features can improve the recognition effect. The purpose of SEBS block is to generate active attention features without any high-level features and enhance the ability of the model to localize lung lesions. In SEBS block, first, a series of candidate templates T with different spatial energy distributions are generated and the controllability of the energy distribution in T enables active attention features to maintain the continuity and integrity of the feature space distributions. Second, Top-ntemplates are selected from T according to certain learning rules, which are then operated by a convolution layer for generating supervision information that can guide the inputs of SEBS block to form active attention features. We evaluated the PKA2-Net on the binary classification problem of identifying pneumonia and healthy controls on a dataset containing 5856 chest X-ray images (ChestXRay2017), the results showed that our method can achieve 97.63% accuracy and 0.9872 sensitivity.

In clinical practice, it is desirable for medical image segmentation models to be able to continually learn on a sequential data stream from multiple sites, rather than a consolidated dataset, due to storage cost and privacy restrictions. However, when learning on a new site, existing methods struggle with a weak memorizability for previous sites with complex shape and semantic information, and a poor explainability for the memory consolidation process. In this work, we propose a novel Shape and Semantics-based Selective Regularization ( [Formula: see text]) method for explainable cross-site continual segmentation to maintain both shape and semantic knowledge of previously learned sites. Specifically, [Formula: see text] method adopts a selective regularization scheme to penalize changes of parameters with high Joint Shape and Semantics-based Importance (JSSI) weights, which are estimated based on the parameter sensitivity to shape properties and reliable semantics of the segmentation object. This helps to prevent the related shape and semantic knowledge from being forgotten. Moreover, we propose an Importance Activation Mapping (IAM) method for memory interpretation, which indicates the spatial support for important parameters to visualize the memorized content. We have extensively evaluated our method on prostate segmentation and optic cup and disc segmentation tasks. Our method outperforms other comparison methods in reducing model forgetting and increasing explainability. Our code is available at https://github.com/jingyzhang/S3R.

In medical image analysis, in order to reduce the impact of unbalanced data sets on data-driven deep learning models, according to the characteristic that the area under the Precision-Recall curve (AUCPR) is sensitive to each category of samples, a novel Harmony loss function with fast convergence speed and high stability was constructed. Since AUCPRneeds to be calculated in discrete domain, in order to ensure the continuous differentiability and gradient existence of the Harmony loss, first, the Logistic function was used to approximate the Logical function in AUCPR. Then, to improve the optimization speed of the Harmony loss during model training, a method of manually setting a certain number of classification thresholds was proposed to further approximate the calculation of AUCPR. After the above two approximate calculation processes, the Harmony loss with stable gradient and high computational efficiency was designed. In the optimization process of the model, since Harmony loss can reconcile recall and precision of each category under different classification thresholds, thereby, it can not only improve the model's ability to recognize categories with less samples, but also maintain the stability of the training curve. To comprehensively evaluate the effects of Harmony loss function, we performed experiments on image 3D reconstruction, 2D segmentation, and unbalanced classification tasks. Experimental results showed that the Harmony loss achieved the state-of-the-art results on four unbalanced data sets. Moreover, the Harmony loss can be easily combined with existing loss functions, and is suitable for most common deep learning models.

Respiratory motion estimation is an important part in image-guided radiation therapy and clinical diagnosis. However, most of the respiratory motion estimation methods rely on indirect measurements of external breathing indicators, which will not only introduce great estimation errors, but also bring invasive injury for patients. In this paper, we propose a method of lung respiratory motion estimation based on fast Kalman filtering and 4D CT image registration (LRME-4DCT). In order to perform dynamic motion estimation for continuous phases, a motion estimation model is constructed by combining two kinds of GPU-accelerated 4D CT image registration methods with fast Kalman filtering method. To address the high computational requirements of 4D CT image sequences, a multi-level processing strategy is adopted in the 4D CT image registration methods, and respiratory motion states are predicted from three independent directions. In the DIR-lab dataset and POPI dataset with 4D CT images, the average target registration error (TRE) of the LRME-4DCT method can reach 0.91 mm and 0.85 mm respectively. Compared with traditional estimation methods based on pair-wise image registration, the proposed LRME-4DCT method can estimate the physiological respiratory motion more accurately and quickly. Our proposed LRME-4DCT method fully meets the practical clinical requirements for rapid dynamic estimation of lung respiratory motion.

To solve the problem that traditional image registration methods based on continuous optimization for large motion lung 4D CT image sequences are easy to fall into local optimal solutions and lead to serious misregistration, a novel image registration method based on high-order Markov Random Field (MRF) is proposed. By analyzing the effect of the deformation field constraint of the potential functions with different order cliques in MRF model, energy functions with high-order cliques form are designed separately for 2D and 3D images to preserve topology of the deformation field. In order to preserve the topology of the deformation field more effectively, it is necessary to apply a smooth term and a topology preservation term simultaneously in the energy function and use logarithmic function to impose a penalty on the Jacobian matrix with high-order cliques in the topology preservation term. For the complexity of the designed energy function with high-order cliques form, Markov Chain Monte Carlo (MCMC) algorithm is used to solve the optimization problem of the designed energy function. To address the high computational requirements in lung 4D CT image registration, a multi-level processing strategy is adopted to reduce the space complexity of the proposed registration method and promotes the computational efficiency. In the DIR-lab dataset with 4D CT images and the COPD (Chronic Obstructive Pulmonary Disease) dataset with 3D CT images, the average target registration error (TRE) of our proposed method can reach 0.95 mm respectively.