上浦 尚武

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.

Surgical skill levels of young ophthalmologists tend to be instinctively judged by ophthalmologists in practice, and hence a stable evaluation is not always made for a single ophthalmologist. Although it has been said that standardizing skill levels presents difficulty as surgical methods vary greatly, approaches based on machine learning seem to be promising for this objective. In this study, we propose a method for displaying the information necessary to quantify the surgical techniques of cataract surgery in real-time. The proposed method consists of two steps. First, we use InceptionV3, an image classification network, to extract important surgical phases and to detect surgical problems. Next, one of the segmentation networks, scSE-FC-DenseNet, is used to detect the cornea and the tip of the surgical instrument and the incisional site in the continuous curvilinear capsulorrhexis, a particularly important phase in cataract surgery. The first and second steps are evaluated in terms of the area under curve (i.e., AUC) of the figure of the true positive rate versus (1—false positive rate) and the intersection over union (i.e., IoU) obtained by the ground truth and prediction associated with the region of interest. As a result, in the first step, the network was able to detect surgical problems with an AUC of 0.97. In the second step, the detection rate of the cornea was 99.7% when the IoU was 0.8 or more, and the detection rates of the tips of the forceps and the incisional site were 86.9% and 94.9% when the IoU was 0.1 or more, respectively. It was thus expected that the proposed method is one of the basic techniques to achieve the standardization of surgical skill levels.

In this paper, a support-vector-machine(SVM)- based method of detecting stenosis is presented for fallopian tubal models. It copes with stenosis detection as classification of data prepared from results of ultrasonic measurements conducted for tubal models. Under assumption that waves reflected at the second and third boundary surfaces of the models potentially include characteristics associated with blocked sections (i.e., stenosis), the method determines the time range having the reflected waves, by referring to maximal values on envelope curves of them The determined range is divided into regular short intervals, and the difference between maximum value and minimum value on envelope curves is calculated for each interval. The ten-dimensional data used to SVM learning and stenosis detection is prepared from the frequency distribution of the number of the short intervals versus difference values. Experimental results establish that the method can achieves favorable accuracies in checking occurrence of stenosis and in identifying tubal model types.

The present study aimed to conduct a real-time automatic analysis of two important surgical phases, which are continuous curvilinear capsulorrhexis (CCC), nuclear extraction, and three other surgical phases of cataract surgery using artificial intelligence technology. A total of 303 cases of cataract surgery registered in the clinical database of the Ophthalmology Department of Tsukazaki Hospital were used as a dataset. Surgical videos were downsampled to a resolution of 299 × 168 at 1 FPS to image each frame. Next, based on the start and end times of each surgical phase recorded by an ophthalmologist, the obtained images were labeled correctly. Using the data, a neural network model, known as InceptionV3, was developed to identify the given surgical phase for each image. Then, the obtained images were processed in chronological order using the neural network model, where the moving average of the output result of five consecutive images was derived. The class with the maximum output value was defined as the surgical phase. For each surgical phase, the time at which a phase was first identified was defined as the start time, and the time at which a phase was last identified was defined as the end time. The performance was evaluated by finding the mean absolute error between the start and end times of each important phase recorded by the ophthalmologist as well as the start and end times determined by the model. The correct response rate of the cataract surgical phase classification was 90.7% for CCC, 94.5% for nuclear extraction, and 97.9% for other phases, with a mean correct response rate of 96.5%. The errors between each phase’s start and end times recorded by the ophthalmologist and those determined by the neural network model were as follows: CCC’s start and end times, 3.34 seconds and 4.43 seconds, respectively and nuclear extraction’s start and end times, 7.21 seconds and 6.04 seconds, respectively, with a mean of 5.25 seconds. The neural network model used in this study was able to perform the classification of the surgical phase by only referring to the last 5 seconds of video images. Therefore, our method has performed like a real-time classification.

In this paper, a method of achieving high accuracy in discriminating right and left eyes is proposed for ophthalmic surgery. A VGG16 convolutional neural network is employed to construct a main classifier. The data presented to the main classifier are some frames sampled at regular intervals from surgery videos. Before classifying the frames to be examined, the proposed method determines whether they are suitable or not to improve the discrimination accuracy as high as possible. In other words, the frames causing erroneous discrimination are omitted. The determination of frames depends on image characteristics associated with lightness values and edges, and on positions of eyelid speculums in them. It is based on SegNet neural networks. The proposed method determines the frames to be presented, if rectangular areas specified by bent parts of the speculums adequately appear in the predefined regions inside them. Experimental results reveal that the proposed method achieves the favorable discrimination accuracy with a small number of data in training SegNet networks compared with another method.

© 2018 IEEE. In this paper, we propose a new approach to classify ovarian follicles into two classes. A smoothing filter which is designed to consider speckle patterns under the resolution of the ultrasound devices is applied for filtering ovarian follicle images. Then, convolutional neural networks are used for extracting features from the filtered ovarian follicle images. Finally, both features extracted by CNNs from the filtered ovarian follicle images and numerical features defined by our previous works are used for classification. From experimental results, we show the effectiveness of our proposed method.

In this paper, a method of determining examinations is proposed for ophthalmologic outpatients, using feedforward neural network (NN for short). The determination is based on the data classification. The proposed method defines four classes of ophthalmologic examinations. It prepares data for NN training and examination determination from handwriting sentences in outpatients' interview sheets. A set of the training data is prepared in the form of a matrix. The words extracted from the sentences are assigned to the matrix columns, while each sheet (or sentences in it) is assigned to a matrix rows. Entries in the matrix takes binary values meaning whether extracted words appear in the sentences in the sheet. The proposed method also the ages of outpatients as entries. NN training is conducted according to the normal backpropagation algorithm using a row as one of the training data. The trained NN has four output neurons each of which takes the value belonging to the range [0, 1]. The class of data to be examined is determined by searching the neuron at which the largest value appears in the output layer. Experimental results comprehensively establish that the proposed method can achieve higher percentages of concordance than other methods.

© 2018 IEEE. In this paper, an ultrasound-measurement-based method of detecting stenosises in fallopian tubal models is presented. The proposed method employs silicone rubber tubes and custom phantoms made of polymer gels as models of fallopian tubes. When the targets are the former (or latter), stenosis models are formed by lumps of polyester sewing threads (or polymer gels with high reflection intensity). The proposed method copes with detecting stenosis models as classification problems by applying support vector machines (SVMs for short). It prepare data for SVM learning and stenosis detection from waveforms in the range specified by two time points at which waves are probably reflected by inner upper and lower surfaces of tubal models. The data preparation is conducted by generating the frequency distribution of the number of regular short intervals into which this range is divided versus normalized values associated with amplitude of the waveforms in such intervals. When it is explored whether given data are prepared from a tube with the stenosis model, a discrimination model constructed by SVM learning works as a two-class classifier. Using the model as a multi-class classifier makes it possible to distinguish data of tubes with stenosis models from each other. It is revealed from experimental results that the proposed method has substantially high capability in detecting stenosis models.

This paper presents a cellular automaton-based model for a mechanical computer called 'Turing Tumble' computer. This computer uses mechanical reactions of a ball flowing down and mechanical components (such as gear and ramp) that are configured on the board, for its computation. A group of cells, called a supercell, is defined in the proposed model in order to represent simultaneous state transition of cells for implementing a chain reaction of connected gears. A small element with its memory called Converter is shown for an illustrative example on this model.

In this paper, a personal identification method is presented for patients visiting departments of ophthalmology. The proposed method prepares data from 32×256-sized matrices, which are inspection results measured by Optical Coherence Tomography, both at the stage of the registration and at that of the collation. It considers a three-tuple with one of top largest Pachymetry values and coordinates specifying its position in the matrix to be a characteristic point, and calculates degree of similarity between any couple of a three-tuple in the registered data and that in the collation data. It employs the sum of the above degrees to determine some subject with registered data as a person to be identified. In addition, it defines some threshold values associated with Instantaneous Keratometric, Reflective Keratometric, and Pachymetry, to eliminate false acceptance as much as possible. Experimental results establish that the proposed method can improve a false acceptance rate while keeping a high identification rate.

© 2017 IEEE. In this paper, a method of calculating inner and outer diameters of fallopian tubes is presented, provided that silicone rubber tubes are models for fallopian tubes. It is based on the ultrasound measurement using a single probe of which the nominal frequency is 5 MHz, and assumes that ultrasound waves emanating from the probe are reflected at the top outer surface, surfaces between water and silicone rubber, and bottom outer surface of the tube. It obtains envelope curves associated with the reflected waves, by applying Hilbert transformation to the waves. It next estimates top four maximal values on the envelope curves to acquire time points when the wave reflections occur. The inner and outer diameters of the tubes are easily estimated by substituting the above time points in simple formulas. Experimental results reveal that the proposed method can achieve substantial accuracy in estimating inner and outer diameters of the target tubes.

In this paper, a method of estimating binding time is proposed for ophthalmological outpatients. Binding time equals the sum of transit time and waiting time at a hospital. It determines the number of outpatients virtually arriving and their arrival time intervals according to the gamma distribution and the exponential distribution. It then assigns examinations to outpatients virtually arriving, referring to their appointment types probabilistically determined. Ten examination rules are established to simulate the passage of time required for examinations and consultation. The proposed method accesses database associated with public transport via some Application Programming Interface, and finally provides transit time and waiting time estimated by the simulation. Experimental results on estimating waiting time show that the average daily difference between actuality and estimation is acceptable.

Copyright © 2017 The Institute of Electronics, Information and Communication Engineers. To promote effective guidance by health checkup results, this paper predict a likelihood of developing lifestyle-related diseases from health check data. In this paper, we focus on the fluctuation of hemoglobin A1c (HbA1c) value, which deeply connected with diabetes onset. Here we predict incensement of HbA1c value and examine which kind of health checkup item has important role for HbA1c fluctuation. Our experimental results show that, when we classify the subjects according to their gender and triglyceride (TG) fluctuation value, we will effectively evaluate the risk of diabetes onset for each class.

To promote effective guidance by health checkup results, this paper predict a likelihood of developing lifestyle-related diseases from health check data. In this paper, we focus on the fluctuation of hemoglobin A1c (HbA1c) value, which deeply connected with diabetes onset. Here we predict incensement of HbA1c value and examine which kind of health checkup item has important role for HbA1c fluctuation. Our experimental results show that, when we classify the subjects according to their gender and triglyceride (TG) fluctuation value, we will effectively evaluate the risk of diabetes onset for each class.

In this paper, we present a method of analyzing relationships between items in specific health examination data, as one of the basic researches to address increases of lifestyle-related diseases. We use self-organizing maps, and pick up the data from the examination dataset according to the condition specified by some item values. We then focus on twelve items such as hemoglobin A1c (HbA1c), aspartate transaminase (AST), alanine transaminase (ALT), gamma-glutamyl transpeptidase (gamma-GTP), and triglyceride (TG). We generate training data presented to a map by calculating the difference between item values associated with successive two years and normalizing the values of this calculation. We label neurons in the map on condition that one of the item values of training data is employed as a parameter. We finally examine the relationships between items by comparing results of labeling (clusters formed in the map) to each other. From experimental results, we separately reveal the relationships among HbA1c, AST, ALT, gamma-GTP and TG in the unfavorable case of HbA1c value increasing and those in the favorable case of HbA1c value decreasing.