Naotake KAMIURA

Abstract Artificial intelligence (AI)-assisted morphological analysis using whole slide images (WSIs) shows promise in supporting complex pathological diagnosis. However, the implementation in clinical settings is costly and demands extensive data storage. This study aimed to develop a compact, practical classification model using patch images selected by pathologists from representative disease areas under a microscope. To evaluate the limits of classification performance, we applied multiple pre-training strategies and convolutional neural networks (CNNs) specifically for the diagnosis of particularly challenging malignant lymphomas and their subtypes. The EfficientNet CNN, pre-trained with ImageNet, exhibited the highest classification performance among the tested models. Our model achieved notable accuracy in a four-class classification (normal lymph node and three B cell lymphoma subtypes) using only hematoxylin and eosin (H&E) stained specimens (AUC = 0.87), comparable to results from immunohistochemical and genetic analyses. This finding suggests that the proposed model enables pathologists to independently prepare image data and easily access the algorithm and enhances diagnostic reliability while significantly reducing costs and time for additional tests, offering a practical and efficient diagnostic support tool for general medical facilities.

PURPOSE: To understand real-world eye drop adherence among glaucoma patients and evaluate the performance of our proposed cloud-based support for eye drop adherence (CASEA). DESIGN: Prospective, observational case series. METHODS: Setting: The Department of Ophthalmology at Tsukazaki Hospital. Patient or study population: Glaucoma patients treated at the hospital from May 2021 to September 2022, with 61 patients initially enrolled. Intervention or observation procedures: Pharmacists guided eye drop administration before the study. Changes in bottle orientation were detected using an accelerometer attached to the container, and acceleration waveforms and date/time data were recorded. Patients visited the clinic during the 4th and 8th weeks to report their eye drop administration, and the data were uploaded to the cloud. Main outcome measures: Two AI models (B-LSTM) were created to analyze the eye drop bottle movement time-series data for patients treating one or both eyes. The models were evaluated by comparing the true administration status with the AI model judgment. RESULTS: Four of the 61 study subjects dropped out. The remaining 57 patients achieved recall, precision, and accuracy values of 98.6 %, 98.6 %, and 95.9 %, respectively, for the two-eyes model and 95.8 %, 98.8 %, and 95.6 % for the one-eye model. Three low-accuracy participants (77.1 %, 71.0 %, and 81.0 %) improved to 100 %, 99.1 %, and 100 %, respectively, after undergoing an additional 8-week performance validation using an aid-type container designed to ensure that the bottle was fully inverted during instillation. CONCLUSIONS: CASEA precisely monitored daily eye drop adherence and enhanced treatment efficacy by identifying patients with difficulty self-medicating. This system has the potential to improve glaucoma patient outcomes by supporting adherence.

In this study, we investigate model structures for neural ODEs to improve the data efficiency in learning the dynamics of control systems. We introduce two model structures and compare them with a typical baseline structure. The first structure considers the relationship between the coordinates and velocities of the control system, while the second structure adds linearity with respect to the control term to the first structure. Both of these structures can be easily implemented without requiring additional computation. In numerical experiments, we evaluate these structure on simulated simple pendulum and CartPole systems and show that incorporating these characteristics into the model structure leads to accurate learning with a smaller amount of training data compared to the baseline structure.