山本 真理子, 吉澤 晋

Mariko YAMAMOTO, Shin YOSHIZAWA

東北大学大学院工学研究科通信工学専攻

Graduate School of Engineering, Tohoku University

キーワード : displacement detection, deep learning, spatial resolution, detection accuracy, high intensity focused ultrasound

目的：超音波診断装置を用いた，2次元変位のサブピクセル精度，高い空間分解能での検出を目的とした．従来の方法は近傍の一様性を仮定しており空間分解能と精度を両立不可能であった．対象と方法：超音波画像を入力，変位場を出力とするDeep Learningネットワークを提案した．オプティカルフロー推定に使用されるFlowNet2を元にネットワーク構造を構築し，学習データをシミュレーションにより作成した．提案方法の変位検出精度と空間分解能をシミュレーションによって評価し，実データでの有用性を強力集束超音波（high intensity focused ultrasound: HIFU）照射下ブタ肝臓の超音波画像によって評価した．これらの結果を従来方法であるLucas-Kanade法と比較した．結果と考察：ピクセルサイズ67 μm四方，信号ノイズ1％，±40 μm以下の変位に対して，ラテラル方向，アキシャル方向それぞれ正確度0.5 μm，0.2 μm以上，精度0.4 μm，0.3 μm以上，空間分解能1.1 mm，0.8 mmが得られた．実験データでも同様の性能向上と，HIFU照射による変性領域の境界が検出可能であることを確認した．結論：Deep Learningに基づく，空間分解能の高いサブピクセル精度の2次元変位検出を実現した．提案方法により，HIFU照射によって生じる組織の微小な変形のモニタリングを可能にした．

Purpose: The purpose of this study was to detect two dimensional and sub-pixel displacement with high spatial resolution using an ultrasonic diagnostic apparatus. Conventional displacement detection methods assume neighborhood uniformity and cannot achieve both high spatial resolution and sub-pixel displacement detection. Subjects and Methods: A deep-learning network that utilizes ultrasound images and output displacement distribution was developed. The network structure was constructed by modifying FlowNet2, a widely used network for optical flow estimation, and a training dataset was developed using ultrasound image simulation. Detection accuracy and spatial resolution were evaluated via simulated ultrasound images, and the clinical usefulness was evaluated with ultrasound images of the liver exposed to high-intensity-focused ultrasound （HIFU）. These results were compared to the Lucas–Kanade method, a conventional sub-pixel displacement detection method. Results and Discussion: For a displacement within ±40 μm （±0.6 pixels）, a pixel size of 67 μm, and signal noise of 1％, the accuracy was above 0.5 μm and 0.2 μm, the precision was above 0.4 μm and 0.3 μm, and the spatial resolution was 1.1 mm and 0.8 mm for the lateral and axial displacements, respectively. These improvements were also observed in the experimental data. Visualization of the lateral displacement distribution, which determines the edge of the treated lesion using HIFU, was also realized. Conclusion: Two-dimensional and sub-pixel displacement detection with high spatial resolution was realized using a deep- learning methodology. The proposed method enabled the monitoring of small and local tissue deformations induced by HIFU exposure.