Early-stage breast cancer is often asymptomatic, highlighting the critical role of computer-aided diagnostic(CAD) systems in mammography screening. While radiologists often refer to bilateral symmetry to identify abnormalities, most existing CAD methods analyze unilateral views or require image registration, which limits their ability to model structural heterogeneity and often introduces distortion. To address this, we propose a registration-free, structure-aware diagnostic framework that integrates bilateral mammography with soft spatial prompting via Vision Transformers (ViT). By directly concatenating bilateral images and introducing a soft attention mask generated from a lightweight segmentation network, our approach enables end-to-end modeling of cross-breast structural differences without the need for region-of-interest extraction. Extensive evaluations on both public and clinical datasets demonstrate that our method consistently outperforms CNN and lightweight Transformer baselines, achieving up to 0.930 accuracy and 0.972 AUC. To our knowledge, this is the first framework to combine bilateral structural modeling and soft guidance in a unified, interpretable, and scalable ViT-based pipeline for breast cancer diagnosis.

Accurate brain tumor segmentation is essential for clinical tasks such as diagnosis, tumor localization, and surgical planning. Although multi-modality MRI provides complementary information about tumor subregions, the difference of image characteristics across modalities poses significant challenges for effective integration. To address this, we proposed a framework for multi-modal brain tumor segmentation that enhances boundary consistency through modality-guided fusion and strengthens tumor discrimination via semantic attention enhancement. Our method incorporates two key modules: an Edge-Enhanced MultiModal Fusion (EMF) module and a Residual Convolutional Block Attention Module (ResCBAM). The EMF module, designed as an early fusion component, leverages 3D Sobel and Laplacian filters to extract structural features and selectively integrates T1c with other modalities to improve boundary-aware representation. At the network bottleneck, ResCBAM combines residual connections with channel and spatial attention to enhance high-level semantic features. Extensive experiments on the BraTS2023 public dataset demonstrate strong segmentation performance, and further evaluation on the external BraTS2025 dataset confirms the robustness of our approach. Ablation studies demonstrate that EMF plays a key role in extracting high-quality, modality-aware representations that can be effectively refined by ResCBAM, highlighting the effectiveness of this cooperative fusion design.

The decline in traditional autopsy practices has led to the rise of autopsy imaging as a non-invasive alternative. However, the shortage of forensic pathologists skilled in postmortem image interpretation presents a significant challenge. Our study addresses this gap by advancing the capabilities of computer-aided diagnosis (CAD) systems in forensic diagnosis. This study builds on a previous work which developed a CAD system for drowning diagnosis and identified a critical limitation: the inconsistency between human expertise and the decision basis of the CAD system. To alleviate this issue, we present an end-to-end CAD system based on Y-Net that not only classifies post-mortem images into drowning or non-drowning but also segments regions of interest in alignment with human expertise. Experiment results showed that the model achieved an accuracy of 0.92 for classification and a mean squared error of 0.04 for segmentation, offering a promising performance and medically consistent results in drowning diagnosis.

Deep learning has proven to be a powerful tool for analyzing complex medical data; however, its effectiveness can be hindered by limited training data, leading to overfitting. In response to this challenge, our paper proposes a novel deep learning-based method that integrates attention optimization to mitigate overfitting when training on few instances. We introduce a unique loss function that not only considers the disparity between predicted scores and class labels but also accounts for the distinction between the model’s attention and human observations. By optimizing for this loss, our model is encouraged to learn essential features identified by experts, enhancing its classification capabilities. To validate the effectiveness of our approach, we focus on the classification of post-mortem computed tomography (PMCT) as a benchmark task, showcasing improved performance even in scenarios with limited training data. Our contributions offer a promising avenue for enhancing the robustness of deep learning models in medical applications with constrained datasets.

It is challenging to diagnose drowning in autopsy even with the help of post-mortem multi-slice computed tomography (MSCT) due to the complex pathophysiology and the shortage of forensic specialists equipped with radiology knowledge. Therefore, a computer-aided diagnosis (CAD) system was developed to help with diagnosis. Most deep learning-based CAD systems only utilize 2D information, which is proper for 2D data such as chest X-ray images. However, 3D information should also be considered for 3D data like CT. Conventional 3D methods require a huge amount of data and computational cost when using 3D methods. In this article, we proposed a 2.5D method that converts 3D data into 2D images to train 2D deep learning models for drowning diagnosis. The key point of this 2.5D method is that it uses a subset to represent the whole case, covering this case as much as possible while avoiding other repetitive information. To evaluate the effectiveness of the proposed method, conventional 2D, previous 2.5D, and 3D deep learning-based methods were tested using an MSCT dataset obtained from Tohoku university. Then, to provide explainable diagnosis results, a visualization method called Gradient-weighted Class Activation Mapping was employed to visualize features relevant to drowning in CT images. Results on drowning diagnosis showed that our proposed method achieved the best performance compared to other 2D, 2.5D, and 3D methods. The visual assessment also demonstrated that our method could find the saliency regions corresponding to drowning.

This letter summarizes and proves the concept of bounded-input bounded-state (BIBS) stability for weight convergence of a broad family of in-parameter-linear nonlinear neural architectures (IPLNAs) as it generally applies to a broad family of incremental gradient learning algorithms. A practical BIBS convergence condition results from the derived proofs for every individual learning point or batches for real-time applications.

This article is to comment on the derivation of the weight-update stability of in-parameter-linear nonlinear learning system with the gradient descent learning rule in the above article. Our comments are not to disqualify the commented article's whole contribution; however, the issues should be pointed out to avoid their proliferation.

Drowning death rate is high in Japan and its diagnosis is still one of the most challenging tasks in the field of forensics due to the complex interpretation of its pathology. Postmortem lungs computed tomography (CT) images can be used for interpretation of forensic pathology due to its benefits but shortage of specialists is a critical problem. Also, manually interpreting CT images is a tiring and time-taking process. In this paper, we proposed a computer-aided diagnosis system based on a deep convolutional neural network (DCNN) for classifying the post-mortem lungs CT images into drowning and non-drowning. A pre-trained DCNN was implemented in this study for classification of post-mortem lungs CT images. The DCNN was trained and tested using a post-mortem lungs CT image database obtained from Tohoku University Autopsy Imaging Center. The training process involves fine-tuning. The experimental results demonstrated a receiver operating characteristic (ROC) curve and an area under the curve (AUC) of 95 percent was achieved in drowning detection using the post-mortem lungs CT images.

The usefulness of taking mammography has widely been recognized, but screening mammography occasionally results in an excessive recommendation for subsequent biopsy causing many women inconvenience and severe anxiety. Especially, there is a high chance of unnecessary biopsy recommendation for those findings which are difficult to be classified into malignancy and benignancy. However, few have focused on the computer-aided diagnosis (CAD) performance for such difficult cases. To address this problem, we developed a deep learning based classification technique to aid the difficult diagnosis. We evaluated 100 benign and malignant masses of the breast imaging-reporting and data system (BI-RADS) Category 4 that are generally difficult to be classified into malignant and benign. Five certificated doctors participated in the experiments where each doctor reads the 100 images alone first and a week later reads again with the proposed CAD system. The area under the receiver operating characteristic curve (AUC-ROC) for the CAD system was 0.79. This is greater than 0.65, the average value of the human readers' AUC-ROCs, while the average value of the human readers' AUC-ROCs reached the best value of 0.8 when they used the CAD system. These results suggest that the proposed CAD system is able to not only outperform human readers in classifying the masses, but also enhance the human performance in this difficult task.

Stability evaluation of a weight-update system of higher order neural units (HONUs) with polynomial aggregation of neural inputs (also known as classes of polynomial neural networks) for adaptation of both feedforward and recurrent HONUs by a gradient descent method is introduced. An essential core of the approach is based on the spectral radius of a weight-update system, and it allows stability monitoring and its maintenance at every adaptation step individually. Assuring the stability of the weight-update system (at every single adaptation step) naturally results in the adaptation stability of the whole neural architecture that adapts to the target data. As an aside, the used approach highlights the fact that the weight optimization of HONU is a linear problem, so the proposed approach can be generally extended to any neural architecture that is linear in its adaptable parameters.

This paper presents a case study of non-Shannon entropy, i.e. Learning Entropy (LE), for instant detection of onset of epileptic seizures in individual EEG time series. Contrary to entropy methods of EEG evaluation that are based on probabilistic computations, we present the LE-based approach that evaluates the conformity of individual samples of data to the contemporary learned governing law of a learning system and thus LE can detect changes of dynamics on individual samples of data. For comparison, the principle and the results are compared to the Sample Entropy approach. The promising results indicate the LE potentials for feature extraction enhancement for early detection of epileptic seizures on individual-data-sample basis.

This paper presents recently introduced concept of Learning Entropy (LE) for time series and recalls the practical form of its evaluation in real time. Then, a technique that estimates the increased risk of prediction inaccuracy of adaptive predictors in real time using LE is introduced. On simulation examples using artificial signal and real respiratory time series, it is shown that LE can be used to evaluate the actual validity of the adaptive predicting model of time series in real time. The introduced technique is discussed as a potential approach to the improvement of accuracy of lung tumor tracking radiation therapy.

A cycling wheel chair (CWC) is a pedal-driven wheelchair developed as a personal transportation device for patients with hemiplegia after cerebral stroke. The CWC enables the patient to travel much faster than an ordinary wheel chair and is effective for prevention of disuse syndrome in lower limbs. However, the patients often have impairment of cognitive function as well as motor function. It is dangerous for such patients to ride on the CWC on outdoor roads. To cope with this problem, a virtual reality system for the CWC named “Virtual CWC” was developed. The Virtual CWC can provide safe and space-saving training and test environment for the patients. In the present study, a new scenario of the Virtual CWC has been developed to test the patient's cognitive function with respect to navigation as well as driving skills. The normal subjects' “homing vectors” obtained after riding on the Virtual CWC in three kinds of routes were analyzed, and compared between a young group (26 normal subjects; age 23.9 ± 2.5) and an elderly group (14 normal subjects; age 69.9 ± 4.1). As a result, the effect of cooperation between pedaling and steering could be found in the angular error of homing vectors in the case of the elderly group.

This paper presents a study of the dynamic (recurrent) quadratic neural unit (QNU) -a class of higher order network or a class of polynomial neural network- as applied to the prediction of lung respiration dynamics. Human lung motion during respiration features nonlinear dynamics and displays quasiperiodical or even chaotic behavior. An attractive approximation capability of the recurrent QNU are demonstrated on a long term prediction of time series generated by chaotic MacKey-Glass equation, by another highly nonlinear periodic time series, and on real lung motion measured during patients respiration. The real time recurrent learning (RTRL) rule is derived for dynamic QNU in a matrix form that is also efficient for implementation. It is shown that the standalone QNU gives promising results on a longer prediction times of the lung position compared to results in recent literature. In the end, we show even more precise results of two QNUs implemented as two local nonlinear predictive models and thus we present and discus a promising direction for high precision prediction of lung motion.

A recurrent neural network is proposed to deal with the convex optimization problem. By employing a specific nonlinear unit, the proposed neural network is proved to be convergent to the optimal solution in finite time, which increases the computation efficiency dramatically. Compared with most of existing stability conditions, i.e., asymptotical stability and exponential stability, the obtained finite-time stability result is more attractive, and therefore could be considered as a useful supplement to the current literature. In addition, a switching structure is suggested to further speed up the neural network convergence. Moreover, by using the penalty function method, the proposed neural network can be extended straightforwardly to solving the constrained optimization problem. Finally, the satisfactory performance of the proposed approach is illustrated by two simulation examples.

In this paper, we propose a new computer aided diagnosis method of pulmonary nodules in X-ray CT images to reduce false positive (FP) rate under high true positive (TP) rate conditions. An essential core of the method is to extract and combine two novel and effective features from the raw CT images: One is orientation features of nodules in a region of interest (ROI) extracted by a Gabor filter, while the other is variation of CT values of the ROI in the direction along body axis. By using the extracted features, a principal component analysis technic and any pattern recognition technics such as neural network approaches can then used to discriminate between nodule and non-nodule images. Simulation results show that discrimination performance using the proposed features is extremely improved compared to that of the conventional method.

In this paper, we analyze neural spike dynamics of a double feedback neural unit (DFNU). An essential emphasis of the analysis is on use of the DFNU's simple formulations that can provide quantitative analytic results. Comparing dynamics of Hodgkin-Huxley model to that of the DFNU, it is shown that dynamics of the DFNU is also physiologically plausible under a condition. The results suggest that high-frequency firings are relatively appropriate for a neural informational carrier due to the reliability and robustness to noisy inputs. To realize such reliable spike communication, we improved the DFNU's performance by using extra noisy inputs with appropriate amplitudes. Simulation studies show that there is optimal region of the amplitude that makes the DFNU possess the noise-enhanced reliable communication ability as similar to stochastic resonance phenomena.

In this paper, a novel analysis technique is applied for investigating human operators' trial and error learning process to control a nonholonomic system, 2-link planer underactuated manipulator (2PUAM). An essential core of the technique is to use a value function of the reinforcement learning scheme for revealing how the operators can find a control strategy. It is an advantage of the proposed technique compared to the others that a transition of the value function may explain the changes of the operators' strategies during the learning process. According to the results of the analysis, the operators tended to explore an objective trajectory first, and then shift to the tracking control of the trajectory. The tracking was accompanied with acceleration to achieve the goal faster. Interestingly, the acceleration disturbs the objective trajectory due to the complex dynamics of the target, and induces another exploration to get better trajectories. The fact that this phase transition structure under unsupervised learning environment is consistent with previously reported results for a supervised case implies that the structure can be a general nature of human learning process.

We propose a self-organizing neural structure with dynamic and spatial changing weights for a feature space representation of concept formation. An essential core of this self-organization is based on an unsupervised learning with incomplete information for the dynamic changing and an extended Hebbian rule for the spatial changing. A concept formation problem requires the neural network to acquire the complete feature space structure of a concept information using an incomplete observation of the concept. The connection structure or self-organizing network can store with the information structure by using the two rules. The Hebbian rule can create a necessary connection corresponding to a feature space substructure of the complete information. On the other hand, unsupervised learning can delete unnecessary connections. Finally concept formation ability of the proposed neural network is proven under some conditions.