Electrocardiogram (ECG) interpretation is essential for diagnosing a wide range of cardiac abnormalities. While deep learning has shown strong potential for automating ECG classification, many existing models rely on large, computationally intensive architectures that hinder practical deployment. In this paper, we present an empirical study of convolutional neural network (CNN) architectures, exploring tradeoffs between diagnostic accuracy and computational efficiency. We benchmark two established baselines: AttiaNet, a compact model composed of sequential temporal and spatial blocks, and DeepResidualCNN, the winning architecture of the 2021 PhysioNet/Computing in Cardiology Challenge. Building on these, we propose three lightweight models: (i) ParallelCNN, which employs dual temporal and spatial branches for parallel pattern extraction; (ii) ParallelCNNew, a variant with symmetric weight initialization for balanced feature learning; and (iii) SimpleNet, a streamlined architecture that jointly processes temporal and spatial dimensions. Our experiments span three publicly available 12-lead ECG datasets from Germany, China, and the United States, covering binary, multiclass, and multilabel classification tasks across diverse patient populations. We further evaluate the impact of integrating low-cost demographic metadata (age and sex) to improve performance with minimal overhead. To ensure fair comparison, we introduce a unified Efficiency Score that integrates model size, inference speed, memory usage, and AUC performance. By balancing diagnostic performance and efficiency, our models offer a scalable and viable foundation for next-generation AI systems in cardiovascular care.

The success of large-scale pretrained models has established fine-tuning as a standard method for achieving significant improvements in downstream tasks. However, fine-tuning the entire parameter set of a pretrained model is costly. Parameter-efficient transfer learning (PETL) has recently emerged as a cost-effective alternative for adapting pretrained models to downstream tasks. Despite its advantages, the increasing model size and input resolution present challenges for PETL, as the training memory consumption is not reduced as effectively as the parameter usage. In this article, we introduce fine-grained prompt tuning plus (FPT+), a PETL method designed for high-resolution medical image classification, which significantly reduces the training memory consumption compared to other PETL methods. FPT+ performs transfer learning by training a lightweight side network and accessing pretrained knowledge from a large pretrained model (LPM) through fine-grained prompts and fusion modules. Specifically, we freeze the LPM of interest and construct a learnable lightweight side network. The frozen LPM processes high-resolution images to extract fine-grained features, while the side network employs corresponding downsampled low-resolution images to minimize memory usage. To enable the side network to leverage pretrained knowledge, we propose fine-grained prompts and fusion modules, which collaborate to summarize information through the LPM's intermediate activations. We evaluate FPT+ on eight medical image datasets of varying sizes, modalities, and complexities. Experimental results demonstrate that FPT+ outperforms other PETL methods, using only 1.03% of the learnable parameters and 3.18% of the memory required for fine-tuning an entire ViT-B model. Our code is available https://github.com/YijinHuang/FPT.

Self-supervised Learning (SSL) has been widely applied to learn image representations through exploiting unlabeled images. However, it has not been fully explored in the medical image analysis field. In this work, Saliency-guided Self-Supervised image Transformer (SSiT) is proposed for Diabetic Retinopathy (DR) grading from fundus images. We novelly introduce saliency maps into SSL, with a goal of guiding self-supervised pre-training with domain-specific prior knowledge. Specifically, two saliency-guided learning tasks are employed in SSiT: 1) Saliency-guided contrastive learning is conducted based on the momentum contrast, wherein fundus images' saliency maps are utilized to remove trivial patches from the input sequences of the momentum-updated key encoder. Thus, the key encoder is constrained to provide target representations focusing on salient regions, guiding the query encoder to capture salient features. 2) The query encoder is trained to predict the saliency segmentation, encouraging the preservation of fine-grained information in the learned representations. To assess our proposed method, four publicly-accessible fundus image datasets are adopted. One dataset is employed for pre-training, while the three others are used to evaluate the pre-trained models' performance on downstream DR grading. The proposed SSiT significantly outperforms other representative state-of-the-art SSL methods on all downstream datasets and under various evaluation settings. For example, SSiT achieves a Kappa score of 81.88% on the DDR dataset under fine-tuning evaluation, outperforming all other ViT-based SSL methods by at least 9.48%.

We propose a novel segmentation approach based on deep 3D convolutional encoder networks with shortcut connections and apply it to the segmentation of multiple sclerosis (MS) lesions in magnetic resonance images. Our model is a neural network that consists of two interconnected pathways, a convolutional pathway, which learns increasingly more abstract and higher-level image features, and a deconvolutional pathway, which predicts the final segmentation at the voxel level. The joint training of the feature extraction and prediction pathways allows for the automatic learning of features at different scales that are optimized for accuracy for any given combination of image types and segmentation task. In addition, shortcut connections between the two pathways allow high- and low-level features to be integrated, which enables the segmentation of lesions across a wide range of sizes. We have evaluated our method on two publicly available data sets (MICCAI 2008 and ISBI 2015 challenges) with the results showing that our method performs comparably to the top-ranked state-of-the-art methods, even when only relatively small data sets are available for training. In addition, we have compared our method with five freely available and widely used MS lesion segmentation methods (EMS, LST-LPA, LST-LGA, Lesion-TOADS, and SLS) on a large data set from an MS clinical trial. The results show that our method consistently outperforms these other methods across a wide range of lesion sizes.

We present a new method for measuring cortical thickness changes on longitudinal magnetic resonance images (MRIs). The method is voxel-based for computational efficiency and sensitivity to subtle changes, but aims for robustness in establishing correspondences by using geometric features that are defined on the cortical skeletons of each scan. In contrast to existing longitudinal methods, our method does not require deformable registration but rather performs cortex-specific, feature-based matching in a confidence-weighted manner, which allows a skeletal point in one scan to be partially matched to multiple points in another scan, thereby enhancing the stability of the matches. Based on comprehensive experiments and statistical analyses on two datasets, our results show that the proposed method demonstrates greater sensitivity to clinically relevant changes than three other state-of-the-art methods and comparable reproducibility.

Deep learning has traditionally been computationally expensive, and advances in training methods have been the prerequisite for improving its efficiency in order to expand its application to a variety of image classification problems. In this letter, we address the problem of efficient training of convolutional deep belief networks by learning the weights in the frequency domain, which eliminates the time-consuming calculation of convolutions. An essential consideration in the design of the algorithm is to minimize the number of transformations to and from frequency space. We have evaluated the running time improvements using two standard benchmark data sets, showing a speed-up of up to 8 times on 2D images and up to 200 times on 3D volumes. Our training algorithm makes training of convolutional deep belief networks on 3D medical images with a resolution of up to 128×128×128 voxels practical, which opens new directions for using deep learning for medical image analysis.

We present a novel 3-D region-based hidden Markov model (rbHMM) for efficient unsupervised 3-D image segmentation. Our contribution is twofold. First, rbHMM employs a more efficient representation of the image data than current state-of-the-art HMM-based approaches that are based on either voxels or rectangular lattices/grids, thus resulting in a faster optimization process. Second, our proposed novel tree-structured parameter estimation algorithm for the rbHMM provides a locally optimal data labeling that is invariant to object rotation, which is a highly valuable property in segmentation tasks, especially in medical imaging where the segmentation results need to be independent of patient positioning in scanners in order to minimize methodological variability in data analysis. We demonstrate the advantages of our proposed technique over grid-based HMMs by validating on synthetic images of geometric shapes as well as both simulated and clinical brain MRI scans. For the geometric shapes data, our method produced consistently accurate segmentation results that were also invariant to object rotation. For the brain MRI data, our white matter and gray matter segmentation resulted in substantially higher robustness and accuracy levels with improved Dice similarity indices of 4.60% (p=0.0022) and 7.71% , respectively.

We present a new algorithm for computing a polygonal surface from a union of balls. The method computes and connects the singular points of a given union of balls in an efficient manner to approximate the boundary. The algorithm uses the dual shape of the balls to give the resulting surface the correct topology. Our method is simple and demonstrated to be robust

We present a new algorithm for simplifying the shape of 3D objects by manipulating their medial axis transform (MAT). From an unorganized set of boundary points, our algorithm computes the MAT, decomposes the axis into parts, then selectively removes a subset of these parts in order to reduce the complexity of the overall shape. The result is simplified MAT that can be used for a variety of shape operations. In addition, a polygonal surface of the resulting shape can be directly generated from the filtered MAT using a robust surface reconstruction method. The algorithm presented is shown to have a number of advantages over other existing approaches.

One well known application area of volume rendering is the reconstruction and visualization of output from medical scanners like computed tomography (CT). 2D greyscale slices produced by these scanners can be reconstructed and displayed onscreen as a 3D model. Volume visualization of medical images must address two important issues. First, it is difficult to segment medical scans into individual materials based only on intensity values. Second, although greyscale images are the normal method for displaying medical volumes, these types of images are not necessarily appropriate for highlighting regions of interest within the volume. Studies of the human visual system have shown that individual intensity values are difficult to detect in a greyscale image. In these situations colour is a more effective visual feature. We addressed both problems during the visualization of CT scans of abdominal aortic aneurysms. We have developed a classification method that empirically segments regions of interest in each of the 2D slices. We use a perceptual colour selection technique to identify each region of interest in both the 2D slices and the 3D reconstructed volumes. The result is a colourized volume that the radiologists are using to rapidly and accurately identify the locations and spatial interactions of different materials from their scans. Our technique is being used in an experimental post operative environment to help to evaluate the results of surgery designed to prevent the rupture of the aneurysm. In the future, we hope to use the technique during the planning of placement of support grafts prior to the actual operation.