Kei Ichiji

Testing potentials of dynamic quadratic neural unit for prediction of lung motion during respiration for tracking radiation therapy

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.

Mass detection using deep convolutional neural network for mammographic computer-aided diagnosis

In recent years, a deep convolutional neural network (DCNN) has attracted great attention due to its outstanding performance in recognition of natural images. However, the DCNN performance for medical image recognition is still uncertain because collecting a large amount of training data is difficult. To solve the problem of the DCNN, we adopt a transfer learning strategy, and demonstrate feasibilities of the DCNN and of the transfer learning strategy for mass detection in mammographic images. We adopt a DCNN architecture that consists of 8 layers with weight, including 5 convolutional layers, and 3 fully-connected layers in this study. We first train the DCNN using about 1.2 million natural images for classification of 1,000 classes. Then, we modify the last fully-connected layer of the DCNN and subsequently train the DCNN using 1,656 regions of interest in mammographic image for two classes classification: mass and normal. The detection test is conducted on 198 mammographic images including 99 mass images and 99 normal images. The experimental results showed that the sensitivity of the mass detection was 89.9 % and the false positive was 19.2 %. These results demonstrated that the DCNN trained by transfer learning strategy has a potential to be a key system for mammographic mass detection computer-aided diagnosis (CAD). In addition, to the best of our knowledge, our study is the first demonstration of the DCNN for mammographic CAD application.

A Fast Neural Network Approach to Predict Lung Tumor Motion during Respiration for Radiation Therapy Applications

During radiotherapy treatment for thoracic and abdomen cancers, for example, lung cancers, respiratory motion moves the target tumor and thus badly affects the accuracy of radiation dose delivery into the target. A real-time image-guided technique can be used to monitor such lung tumor motion for accurate dose delivery, but the system latency up to several hundred milliseconds for repositioning the radiation beam also affects the accuracy. In order to compensate the latency, neural network prediction technique with real-time retraining can be used. We have investigated real-time prediction of 3D time series of lung tumor motion on a classical linear model, perceptron model, and on a class of higher-order neural network model that has more attractive attributes regarding its optimization convergence and computational efficiency. The implemented static feed-forward neural architectures are compared when using gradient descent adaptation and primarily the Levenberg-Marquardt batch algorithm as the ones of the most common and most comprehensible learning algorithms. The proposed technique resulted in fast real-time retraining, so the total computational time on a PC platform was equal to or even less than the real treatment time. For one-second prediction horizon, the proposed techniques achieved accuracy less than one millimeter of 3D mean absolute error in one hundred seconds of total treatment time.

A kernel-based method for markerless tumor tracking in kV fluoroscopic images

Markerless tracking of respiration-induced tumor motion in kilo-voltage (kV) fluoroscopic image sequence is still a challenging task in real time image-guided radiation therapy (IGRT). Most of existing markerless tracking methods are based on a template matching technique or its extensions that are frequently sensitive to non-rigid tumor deformation and involve expensive computation. This paper presents a kernel-based method that is capable of tracking tumor motion in kV fluoroscopic image sequence with robust performance and low computational cost. The proposed tracking system consists of the following three steps. To enhance the contrast of kV fluoroscopic image, we firstly utilize a histogram equalization to transform the intensities of original images to a wider dynamical intensity range. A tumor target in the first frame is then represented by using a histogram-based feature vector. Subsequently, the target tracking is then formulated by maximizing a Bhattacharyya coefficient that measures the similarity between the tumor target and its candidates in the subsequent frames. The numerical solution for maximizing the Bhattacharyya coefficient is performed by a mean-shift algorithm. The proposed method was evaluated by using four clinical kV fluoroscopic image sequences. For comparison, we also implement four conventional template matching-based methods and compare their performance with our proposed method in terms of the tracking accuracy and computational cost. Experimental results demonstrated that the proposed method is superior to conventional template matching-based methods.

SU‐HH‐BRB‐10: Adaptive Seasonal Autoregressive Model Based Intrafractional Lung Tumor Motion Prediction for Continuously Irradiation

Purpose: To irradiate continuously to a moving tumor, core techniques are to observe position and shape of the target tumor and to adapt the radiation beam to the intra‐fractional motion and deformation. In addition, we need a compensation technique because measurement of the tumor position and control of the radiation device have some time delays. In this study, we propose a new prediction method of lungtumor motion, to compensate the time delays. Method and Materials: An essential core of the proposed method is adaptation to time‐variant nature of lungtumor motion. Lungtumor motion observed at Hokkaido university hospital was used for development of the proposed method. The motion has time‐variant, but periodic nature, that is, the cyclic period changes with time. This nature often causes the rise of the prediction error when we use conventional prediction method for periodical time series (e.g. seasonal autoregressive integral moving‐average model: SARIMA). The proposed method is based on SARIMA model, but was developed to take into account the quasi‐periodic nature of the lungtumor motion. To estimate the time‐variant period, we adopted correlation analysis. Then, the conventional SARIMA model was modified to a time‐variant SARIMA model by using the estimated period. Results: Prediction error of the proposed method was compared with that of the conventional methods, by using real lungtumor motion. Experimental results show that the prediction error of the proposed method was the least. The average of prediction error are 0.7911 [mm] at 0.5[sec] ahead and 0.8818[mm] at 1.0[sec] ahead, respectively. Conclusion: We have developed the new prediction method of the lungtumor motion for compensation of time‐delays of radiation device. The proposed method achieved highly accurate prediction of the real lungtumor motion. The method can thus sufficient for continuously irradiation to the moving lungtumor.

Markerless Lung Tumor Motion Tracking by Dynamic Decomposition of X-Ray Image Intensity

We propose a new markerless tracking technique of lung tumor motion by using an X-ray fluoroscopic image sequence for real-time image-guided radiation therapy (IGRT). A core innovation of the new technique is to extract a moving tumor intensity component from the fluoroscopic image intensity. The fluoroscopic intensity is the superimposition of intensity components of all the structures passed through by the X-ray. The tumor can then be extracted by decomposing the fluoroscopic intensity into the tumor intensity component and the others. The decomposition problem for more than two structures is ill posed, but it can be transformed into a well-posed one by temporally accumulating constraints that must be satisfied by the decomposed moving tumor component and the rest of the intensity components. The extracted tumor image can then be used to achieve accurate tumor motion tracking without implanted markers that are widely used in the current tracking techniques. The performance evaluation showed that the extraction error was sufficiently small and the extracted tumor tracking achieved a high and sufficient accuracy less than 1 mm for clinical datasets. These results clearly demonstrate the usefulness of the proposed method for markerless tumor motion tracking.

A Time-Varying Seasonal Autoregressive Model-Based Prediction of Respiratory Motion for Tumor following Radiotherapy

To achieve a better therapeutic effect and suppress side effects for lung cancer treatments, latency involved in current radiotherapy devices is aimed to be compensated for improving accuracy of continuous (not gating) irradiation to a respiratory moving tumor. A novel prediction method of lung tumor motion is developed for compensating the latency. An essential core of the method is to extract information valuable for the prediction, that is, the periodic nature inherent in respiratory motion. A seasonal autoregressive model useful to represent periodic motion has been extended to take into account the fluctuation of periodic nature in respiratory motion. The extended model estimates the fluctuation by using a correlation-based analysis for adaptation. The prediction performance of the proposed method was evaluated by using data sets of actual tumor motion and compared with those of the state-of-the-art methods. The proposed method demonstrated a high performance within submillimeter accuracy. That is, the average error of 1.0 s ahead predictions was 0.931 ± 0.055 mm. The accuracy achieved by the proposed method was the best among those by the others. The results suggest that the method can compensate the latency with sufficient accuracy for clinical use and contribute to improve the irradiation accuracy to the moving tumor.

Study of Learning Entropy for Novelty Detection in lung tumor motion prediction for target tracking radiation therapy

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.

Respiratory motion prediction for tumor following radiotherapy by using time-variant seasonal autoregressive techniques

We develop a new prediction method of respiratory motion for accurate dynamic radiotherapy, called tumor following radiotherapy. The method is based on a time-variant seasonal autoregressive (TVSAR) model and extended to further capture time-variant and complex nature of various respiratory patterns. The extended TVSAR can represent not only the conventional quasi-periodical nature, but also the residual components, which cannot be expressed by the quasi-periodical model. Then, the residuals are adaptively predicted by using another autoregressive model. The proposed method was tested on 105 clinical data sets of tumor motion. The average errors were 1.28 ± 0.87 mm and 1.75 ± 1.13 mm for 0.5 s and 1.0 s ahead prediction, respectively. The results demonstrate that the proposed method can outperform the state-of-the-art prediction methods.

A Respiratory Motion Prediction Based on Time-Variant Seasonal Autoregressive Model for Real-Time Image-Guided Radiotherapy

In radiation therapy, to deliver continuously a sufficient radiation dose to target volume yields a better therapeutic effect. While, avoiding an exposure to healthy tissues surrounding the target volume is also an important requirement for suppressing the adverse effect. Image-guided radiation therapy (IGRT) has potential to achieve the two requirements and as it’s application, stereotactic body radiation therapy (SBRT) has been used in clinic. In SBRT, the irradiated field is positioned with millimeter accuracy by proper daily setup. The accurate irradiation can allow the increase of radiation dose by ignoring the irradiation to the healthy tissues. Indeed, it has been reported that the treatment result of SBRT is comparable to the outcome from surgery [1].