This paper presents a classification model for eye diseases utilizing attention mechanisms to learn features from fundus images and structures. The study focuses on diagnosing glaucoma by extracting retinal vessels and the optic disc from fundus images using a ResU-Net-based segmentation model and Hough Circle Transform, respectively. The extracted structures and preprocessed images were inputted into a CNN-based multi-input model for training. Comparative evaluations demonstrated that our model outperformed other research models in classifying glaucoma, even with a smaller dataset. Ablation studies confirmed that using attention mechanisms to learn fundus structures significantly enhanced performance. The study also highlighted the challenges in normal case classification due to potential feature degradation during structure extraction. Future research will focus on incorporating additional fundus structures such as the macula, refining extraction algorithms, and expanding the types of classified eye diseases.

Background: In this study, we explored the potential of predicting dysglycemia in patients who need to continuously manage blood glucose levels using a non-invasive method via electrocardiography (ECG). Methods: The data were collected from patients with diabetes, and heart rate variability (HRV) features were extracted via ECG processing. A residual block-based one-dimensional convolution neural network model was used to predict dysglycemia. Results: The dysglycemia prediction results at each time point, including at the time of blood glucose measurement, 15 min prior to measurement, and 30 min prior to measurement, exhibited no significant differences compared with the blood glucose measurement values. This result confirmed that the proposed artificial intelligence model for dysglycemia prediction performed well at each time point. Additionally, to determine the optimal number of features required for predicting dysglycemia, 77 HRV features were individually eliminated in the order of decreasing importance with respect to the prediction accuracy; the optimal number of features for the model to predict dysglycemia was determined to be 12. The dysglycemia prediction results obtained 30 min prior to measurement, which exhibited the highest prediction range in this study, were as follows: accuracy = 90.5, sensitivity = 87.52, specificity = 92.74, and precision = 89.86. Conclusions: Furthermore, we determined that no significant differences exist in the blood glucose prediction results reported in previous studies, wherein various vital signs and blood glucose values were used as model inputs, and the results obtained in this study, wherein only ECG data were used to predict dysglycemia.

Background/Objectives: Multiple instance learning (MIL) is widely used for automated classification of epithelial ovarian cancer subtypes from whole slide images (WSIs), but the relative contributions of feature extractor, aggregation strategy, and classification framework (flat vs. hierarchical) choices remain unclear under severe class imbalance. Methods: We evaluated 36 configurations on 510 WSIs from the UBC-OCEAN dataset using stratified five-fold cross-validation, comparing three pathology foundation models (Phikon-v2, CTransPath, UNI), six aggregators (mean/max pooling, ABMIL, CLAM-SB, DSMIL, DTP-TransMIL), and two classification strategies. Pathologist-annotated WSIs assessed attention map interpretability. Results: Feature extractor selection contributed substantially more variance than aggregator choice. Cascade balanced accuracy ranged from 0.538 (Phikon-v2) to 0.925 (UNI); CTransPath (~32 K pretraining WSIs) reached 0.870, exceeding Phikon-v2 (~58 K WSIs) and approaching UNI (~100 K+ WSIs), indicating that pretraining objective and architecture contribute as substantially as scale. The hierarchical cascade consistently improved high-grade serous carcinoma (HGSC) recall across all six evaluated configurations (+0.073 to +0.530), detecting 206 of 217 cases (0.949) with UNI max pooling. Quantitative spatial alignment analysis confirmed that both stronger feature extractors—CTransPath and UNI—generated significantly more spatially structured attention distributions than Phikon-v2 (paired Wilcoxon, p = 0.008 and p = 0.032, respectively). Conclusions: Feature extractor choice contributed more variance than aggregator selection, with the largest gap between Phikon-v2 and stronger extractors. Hierarchical cascades consistently improved HGSC recall across all configurations.

(1) Background: The misuse of transformation technology using medical images is a critical problem that can endanger patients’ lives, and detecting manipulation via a deep learning model is essential to address issues of manipulated medical images that may arise in the healthcare field. (2) Methods: The dataset was divided into a real fundus dataset and a manipulated dataset. The fundus image manipulation detection model uses a deep learning model based on a Convolution Neural Network (CNN) structure that applies a concatenate operation for fast computation speed and reduced loss of input image weights. (3) Results: For real data, the model achieved an average sensitivity of 0.98, precision of 1.00, F1-score of 0.99, and AUC of 0.988. For manipulated data, the model recorded sensitivity of 1.00, precision of 0.84, F1-score of 0.92, and AUC of 0.988. Comparatively, five ophthalmologists achieved lower average scores on manipulated data: sensitivity of 0.71, precision of 0.61, F1-score of 0.65, and AUC of 0.822. (4) Conclusions: This study presents the possibility of addressing and preventing problems caused by manipulated medical images in the healthcare field. The proposed approach for detecting manipulated fundus images through a deep learning model demonstrates higher performance than that of ophthalmologists, making it an effective method.

(1) Background: The misuse of transformation technology using medical images is a critical problem that can endanger patients’ lives, and detecting manipulation via a deep learning model is essential to address issues of manipulated medical images that may arise in the healthcare field. (2) Methods: The dataset was divided into a real fundus dataset and a manipulated dataset. The fundus image manipulation detection model uses a deep learning model based on a Convolution Neural Network (CNN) structure that applies a concatenate operation for fast computation speed and reduced loss of input image weights. (3) Results: According to the fundus image manipulation detection model’s results, for the real data, the four lesions had average sensitivity = 0.98, precision = 1.00, and F1-score = 0.99. For the manipulated data, the average sensitivity = 1.00, precision = 0.84, and F1-score = 0.92. The average Area Under the Curve (AUC) for the four lesions was 0.988, which is relatively high. Subsequently, to confirm whether the fundus image manipulation detection model can operate effectively in actual clinical environments, the model’s outcomes were compared with those of five ophthalmologists. For real data, the four lesions showed an average sensitivity = 0.93, precision = 0.96, and F1-score = 0.95, whereas for manipulated data, the average sensitivity = 0.71, precision = 0.61, and F1-score = 0.65. The four lesions also had an average AUC = 0.822, which was lower than the results of the fundus image manipulation detection model across all metrics. (4) Conclusions: This study presents the possibility of addressing and preventing problems caused by manipulated medical images in the healthcare field. The proposed approach for detecting manipulated fundus images through a deep learning model demonstrates higher performance than that of ophthalmologists, making it an effective method.

Multimodal studies of prostate cancer typically report aggregate fusion gains from histology and transcriptomics but rarely characterize when each modality is informative. We asked whether morphology and transcriptomics contribute differently to distinct clinical endpoints, using Gleason grading and nodal status prediction. On 401 TCGA-PRAD patients with matched whole-slide images, bulk RNA sequencing, and clinical data, we evaluated 45 morphological configurations, seven RNA configurations, and six fusion strategies across four grading formulations, T-stage, and N-stage. To avoid per-task model selection, morphology used a single fixed configuration (gated ABMIL on UNI features). Contribution estimates used repeated stratified five-fold cross-validation, with a locked conformal-prediction protocol and gene set enrichment analysis. The two endpoints showed opposite modality dependence. For Gleason grade, the fixed morphological model exceeded the best RNA configuration across all 45 configurations (five-class macro-F1: 0.445 versus 0.398). For nodal status, no morphological configuration exceeded AUROC 0.62, whereas transcriptomics reached 0.68 (permutation p = 0.037); a direct interaction test confirmed the reversal (bootstrap 95% CI excluding zero). Fusion gains were modest and task-dependent; enrichment analysis linked the nodal signal to loss of smooth muscle programs, consistent with established dedifferentiation biology. The informative modality is task-dependent: morphology predicts grade, transcriptomics predicts nodal status.