diff --git a/Reference/main.pdf b/Reference/main.pdf new file mode 100644 index 0000000..93df294 --- /dev/null +++ b/Reference/main.pdf Binary files differ diff --git a/Reference/main.tex b/Reference/main.tex new file mode 100644 index 0000000..5953b23 --- /dev/null +++ b/Reference/main.tex @@ -0,0 +1,166 @@ +\documentclass{midl} % Include author names + +% The following packages will be automatically loaded: +% jmlr, amsmath, amssymb, natbib, graphicx, url, algorithm2e +% ifoddpage, relsize and probably more +% make sure they are installed with your latex distribution + +\usepackage{makecell} + +% Header for extended abstracts +\jmlrproceedings{MIDL}{Medical Imaging with Deep Learning} +\jmlrpages{} +\jmlryear{2026} + +% to be uncommented for submissions under review +\jmlrworkshop{Short Paper Track} +\jmlrvolume{} +\editors{} + +\title[CNN-LSTM with Geometric Transformation and Clinical Knowledge Integration]{CNN-LSTM with Geometric Transformation and Clinical Knowledge Integration for No-Reflow and Slow Flow Prediction from Intravascular Ultrasound Images} + + % Use \Name{Author Name} to specify the name. + % If the surname contains spaces, enclose the surname + % in braces, e.g. \Name{John {Smith Jones}} similarly + % if the name has a "von" part, e.g \Name{Jane {de Winter}}. + % If the first letter in the forenames is a diacritic + % enclose the diacritic in braces, e.g. \Name{{\'E}louise Smith} + + % Two authors with the same address + % \midlauthor{\Name{Author Name1} \Email{abc@sample.edu}\and + % \Name{Author Name2} \Email{xyz@sample.edu}\\ + % \addr Address} + + % Three or more authors with the same address: + % \midlauthor{\Name{Author Name1} \Email{an1@sample.edu}\\ + % \Name{Author Name2} \Email{an2@sample.edu}\\ + % \Name{Author Name3} \Email{an3@sample.edu}\\ + % \addr Address} + + +% Authors with different addresses: +% \midlauthor{\Name{Author Name1} \Email{abc@sample.edu}\\ +% \addr Address 1 +% \AND +% \Name{Author Name2} \Email{xyz@sample.edu}\\ +% \addr Address 2 +% } + +%\footnotetext[1]{Contributed equally} + +% More complicate cases, e.g. with dual affiliations and joint authorship +\midlauthor{ + \Name{Keishi Kaneko\nametag{$^{1}$}} \Email{kaneko.keishi@chiba-u.jp}\\ + \addr $^{1}$ Department of Medical Engineering, Graduate School of Science and Engineering, Chiba University, Chiba, Japan \AND + \Name{Takeshi Nishi\nametag{$^{2}$}} \Email{takeshi24@hotmail.co.jp}\\ + \addr $^{2}$ Department of Cardiology, Beacon Kalamazoo Hospital, 1521 Gull Road, Kalamazoo, MI 49048, USA \AND + \Name{Yukihiro Nomura\nametag{$^{3}$}} \Email{ynomura@chiba-u.jp}\\ + \Name{Toshiya Nakaguchi\nametag{$^{3\dagger}$}} \Email{nakaguchi@faculty.chiba-u.jp}\\ + \addr $^{3}$ Center for Frontier Medical Engineering, Chiba University, Chiba, Japan +} + +\begin{document} + +\maketitle +{\renewcommand{\thefootnote}{\dag}\footnotetext{Corresponding author}} + +\begin{abstract} + During percutaneous coronary intervention (PCI), embolized plaque can cause complications + such as no-reflow and slow flow phenomena. + Predicting these complications is clinically crucial but challenging due to a complex interplay of various factors, + such as plaque composition and vessel diameter. + While intravascular ultrasound (IVUS) provides valuable lesion information, + models using only IVUS images as input yield limited predictive performance. + In this study, we propose an approach that integrates geometric transformation and clinical knowledge into a deep learning pipeline + to improve complication prediction. + Specifically, we apply a polar coordinate transformation to IVUS images to improve spatial feature extraction efficiency and eliminate non-region-of-interest artifacts such as the IVUS catheter. + Furthermore, we compute four statistics of the attenuation angle within the plaque across multiple frames---mean, maximum, minimum, and standard deviation---and integrate them into a convolutional neural network--long short-term memory (CNN-LSTM) architecture. + Experimental results on a dataset of 312 cases demonstrate that our proposed integration of polar transformation and clinical features improves the area under the curve (AUC) from 0.62 to 0.66, + indicating its potential to enable more accurate quantitative risk assessment. +\end{abstract} + +\begin{keywords} + Intravascular Ultrasound, Attenuated Plaque, CNN-LSTM +\end{keywords} + +\section{Introduction} +Percutaneous coronary intervention (PCI) is a standard treatment for coronary artery disease; +however, plaque embolization can cause complications such as the no-reflow phenomenon and slow flow, +which occur postoperatively in approximately 10--20\% of cases \citep{babapoor2022}. +These complications, once they occur, require post-procedural treatment such as vasodilators, +which carries the risk of serious adverse effects including acute hypotension and arrhythmias. +Therefore, establishing preventive strategies to avoid these complications before they occur, +rather than managing them after onset, is of critical importance. +In current clinical practice, distal protection devices are effective preventive measures for high-risk lesions; however, their routine application to all cases is impractical due to the burden of device preparation, making accurate pre-procedural risk stratification essential. +However, intravascular ultrasound (IVUS) interpretation requires specialized expertise, and accurate prediction by visual assessment alone is challenging because multiple factors---such as vessel diameter and plaque composition---interact in a complex manner, necessitating methods that quantitatively assess complication risk from IVUS images. +Our preliminary study used an approach in which IVUS images were input directly into a ResNet-34-long short-term memory (LSTM) model and achieved an area under the curve (AUC)---a threshold-independent metric---of only 0.62, indicating that predictive performance requires improvement. +A significant association between attenuated plaque and no-reflow has been reported \citep{wu2011}, +but their scoring system is manual, motivating an automated and objective approach. +In this study, we propose a method that explicitly integrates attenuation angle statistics computed from physician annotations into a convolutional neural network--LSTM (CNN-LSTM) architecture. + +\section{Methods} + +\subsection{Dataset} +The IVUS data used in this study were obtained from IVUS videos acquired at Chiba University Hospital. +For each case, 13 frames were extracted at 1\,mm intervals, centered on the frame corresponding to the plaque lesion, with six frames on either side. +The dataset comprises 312 cases: 39 with no-reflow, 35 with slow flow, and 238 without complications. No-reflow and slow flow cases are combined as the positive class. +Each frame is annotated with the attenuation angle direction by a physician experienced in IVUS image interpretation. +All experiments were evaluated using 5-fold cross-validation, and results are reported as the mean across folds. + +\subsection{Proposed Method} + +First, a polar coordinate transformation is applied to each IVUS frame. +The transformation origin is set to the image center and resamples the image to $224 \times 254$ pixels using bicubic interpolation. +The leftmost 30 pixels, corresponding to the IVUS catheter region, are then removed, +yielding a final input size of $224 \times 224$ pixels. +This preprocessing eliminates non-region-of-interest artifacts and unfolds the circular vessel cross-section +into a rectangular representation, improving spatial feature extraction efficiency. +Next, the attenuation angle annotations are used to compute four statistics across the 13 frames---mean, maximum, minimum, and standard deviation---forming the attenuated plaque feature vector. +The model consists of ResNet-34 and an LSTM; the attenuated plaque feature vector is concatenated +to the LSTM output and passed through a fully connected layer to predict the probability of complication occurrence. +We also evaluate a comparison method using a Region of Interest (ROI) mask of the plaque region concatenated to the IVUS image along the channel dimension, where physician annotations are used for annotated frames and the segmentation model of \citet{nishi2021} fine-tuned on annotated frames of this dataset is used for frames in which severe attenuation makes annotation infeasible. + + +\section{Results} + +\tableref{tab:results} presents the ablation study results, including comparisons with the scoring-based approach of \citet{wu2011} (AUC: 0.65) and a single physician's prediction made by reviewing IVUS images from the same dataset (AUC: 0.56). +The proposed method (AUC: 0.66) outperformed \citet{wu2011}'s approach, indicating that image features extracted from IVUS by the CNN provide additional predictive value beyond the attenuation angle alone. +All AI models also outperformed the physician in AUC. +While polar transformation alone did not improve AUC over the baseline (0.62), integrating attenuated plaque features alone raised it to 0.63. +The proposed method combining both achieved the highest AUC of 0.66, with reduced variance ($\pm 0.07$) compared with the baseline ($\pm 0.12$), suggesting greater stability. +In contrast, adding the ROI mask to the proposed method yielded an AUC of only 0.63, indicating that the ROI mask does not necessarily contribute to improved predictive performance. + +\begin{table}[htbp] + \floatconts + {tab:results}% + {\caption{Ablation study results.}}% + {\resizebox{\linewidth}{!}{\small\begin{tabular}{lccccccc} + \hline + Method & \makecell{Polar \\Transform} & \makecell{Attenuated\\Plaque Features} & Accuracy & Precision & Recall & F1 & AUC \\ + \hline + \citet{wu2011} & --- & --- & $0.58 \pm 0.16$ & $0.31 \pm 0.06$ & $0.54 \pm 0.30$ & $0.36 \pm 0.10$ & $0.65 \pm 0.08$ \\ + Physician prediction & --- & --- & $0.56 \pm 0.07$ & $0.32 \pm 0.05$ & $\mathbf{0.72} \pm 0.08$ & $\mathbf{0.44} \pm 0.06$ & $0.65 \pm 0.09$ \\ + ResNet-34-LSTM & $\times$ & $\times$ & $0.54 \pm 0.12$ & $0.31 \pm 0.08$ & $0.71 \pm 0.18$ & $0.43 \pm 0.09$ & $0.62 \pm 0.12$ \\ + ResNet-34-LSTM & $\checkmark$ & $\times$ & $0.64 \pm 0.13$ & $\mathbf{0.36} \pm 0.14$ & $0.44 \pm 0.09$ & $0.38 \pm 0.11$ & $0.62 \pm 0.12$ \\ + ResNet-34-LSTM & $\times$ & $\checkmark$ & $0.56 \pm 0.11$ & $0.31 \pm 0.07$ & $0.64 \pm 0.20$ & $0.40 \pm 0.05$ & $0.63 \pm 0.09$ \\ + ResNet-34-LSTM (Ours) & $\checkmark$ & $\checkmark$ & $0.60 \pm 0.11$ & $0.35 \pm 0.10$ & $0.62 \pm 0.11$ & $0.43 \pm 0.05$ & $\mathbf{0.66} \pm 0.07$ \\ + ResNet-34-LSTM (+ROI Mask) & $\checkmark$ & $\checkmark$ & $\mathbf{0.65} \pm 0.08$ & $0.35 \pm 0.08$ & $0.53 \pm 0.22$ & $0.41 \pm 0.09$ & $0.63 \pm 0.11$ \\ + \hline + \end{tabular}}} +\end{table} + + +\section{Conclusion} +% 0.25ページ +In this study, we proposed a CNN-LSTM pipeline integrating polar coordinate transformation and attenuated plaque features for predicting PCI complications from IVUS images. +The ablation study demonstrated that combining both components improved the AUC from the baseline of 0.62 to 0.66, outperforming physician prediction in AUC (AUC: 0.56). +These results indicate that explicitly integrating geometric transformation of IVUS images with clinical knowledge is effective in improving the predictive performance of deep learning models for complication prediction. +The failure of the ROI mask to improve predictive performance is likely attributable to insufficient accuracy of plaque region segmentation. +Future work includes further improving predictive performance through the integration of other lesion features, such as calcified plaque, as well as clinical information. + +% Acknowledgments---Will not appear in anonymized version +\midlacknowledgments{This study was approved by the Ethics Committee of the Graduate School of Medicine, Chiba University (Approval No. 3270).} + +\bibliography{references} + +\end{document} \ No newline at end of file diff --git a/Reference/main_ja.pdf b/Reference/main_ja.pdf new file mode 100644 index 0000000..cb6bb1a --- /dev/null +++ b/Reference/main_ja.pdf Binary files differ diff --git a/Reference/main_ja.tex b/Reference/main_ja.tex new file mode 100644 index 0000000..ca25c99 --- /dev/null +++ b/Reference/main_ja.tex @@ -0,0 +1,178 @@ +% !TEX program = lualatex +\documentclass{midl} % Include author names + +% The following packages will be automatically loaded: +% jmlr, amsmath, amssymb, natbib, graphicx, url, algorithm2e +% ifoddpage, relsize and probably more +% make sure they are installed with your latex distribution + +% ---- 日本語サポート (LuaLaTeX用) ---- +\usepackage{luatexja} +\usepackage{luatexja-fontspec} + +\usepackage{makecell} + +% Header for extended abstracts +\jmlrproceedings{MIDL}{Medical Imaging with Deep Learning} +\jmlrpages{} +\jmlryear{2026} + +% to be uncommented for submissions under review +\jmlrworkshop{Short Paper -- MIDL 2026 submission} +\jmlrvolume{-- Under Review} +\editors{Under Review for MIDL 2026} + +\title[CNN-LSTM with Geometric Transformation and Clinical Knowledge Integration]{CNN-LSTM with Geometric Transformation and Clinical Knowledge Integration for No-Reflow and Slow Flow Prediction from Intravascular Ultrasound Images} + + % Use \Name{Author Name} to specify the name. + % If the surname contains spaces, enclose the surname + % in braces, e.g. \Name{John {Smith Jones}} similarly + % if the name has a "von" part, e.g \Name{Jane {de Winter}}. + % If the first letter in the forenames is a diacritic + % enclose the diacritic in braces, e.g. \Name{{\'E}louise Smith} + + % Two authors with the same address + % \midlauthor{\Name{Author Name1} \Email{abc@sample.edu}\and + % \Name{Author Name2} \Email{xyz@sample.edu}\\ + % \addr Address} + + % Three or more authors with the same address: + % \midlauthor{\Name{Author Name1} \Email{an1@sample.edu}\\ + % \Name{Author Name2} \Email{an2@sample.edu}\\ + % \Name{Author Name3} \Email{an3@sample.edu}\\ + % \addr Address} + + +% Authors with different addresses: +% \midlauthor{\Name{Author Name1} \Email{abc@sample.edu}\\ +% \addr Address 1 +% \AND +% \Name{Author Name2} \Email{xyz@sample.edu}\\ +% \addr Address 2 +% } + +%\footnotetext[1]{Contributed equally} + +% More complicate cases, e.g. with dual affiliations and joint authorship +\midlauthor{ + \Name{Keishi Kaneko\nametag{$^{1}$}} \Email{kaneko.keishi@chiba-u.jp}\\ + \addr $^{1}$ Department of Medical Engineering, Graduate School of Science and Engineering, Chiba University, Chiba, Japan \AND + \Name{Takeshi Nishi\nametag{$^{2}$}} \Email{takeshi24@hotmail.co.jp}\\ + \addr $^{2}$ Department of Cardiology, Beacon Kalamazoo Hospital, 1521 Gull Road, Kalamazoo, MI 49048, USA \AND + \Name{Yukihiro Nomura\nametag{$^{3}$}} \Email{ynomura@chiba-u.jp}\\ + \Name{Toshiya Nakaguchi\nametag{$^{3\dagger}$}} \Email{nakaguchi@faculty.chiba-u.jp}\\ + \addr $^{3}$ Center for Frontier Medical Engineering, Chiba University, Chiba, Japan +} + +\begin{document} + +\maketitle +{\renewcommand{\thefootnote}{\dag}\footnotetext{Corresponding author}} + +\begin{abstract} + 経皮的冠動脈形成術（PCI）において，流出したプラークが末梢血管を完全に閉塞するNo-reflow現象や， + 血流が著しく低下するSlow flow現象といった合併症を予測することは臨床的に重要である． + しかし，これらの合併症の発症にはプラーク組成や血管径など様々な要因が複合的に絡むため， + 定量的なリスク評価は困難である． + 血管内超音波（IVUS）画像は病変部の有用な情報を提供するものの，IVUS画像のみを入力する従来の手法では， + 予測性能に限界があった． + 本研究では，深層学習パイプラインに幾何学変換と医学的知見を統合することで，これら合併症の発症予測精度を向上させる手法を提案する． + 具体的には，IVUS画像に極座標変換を適用し，IVUSカテーテルなどの非関心領域を除去し，血管構造の特徴抽出の効率化を図る． + さらに，複数フレームにわたってプラーク内の減衰角度の統計量---平均・最大値・最小値・標準偏差の4統計量---を算出し，これをconvolutional neural network--long short-term memory（CNN-LSTM）アーキテクチャに統合した． + 312症例を対象とした実験の結果，極座標変換と減衰プラーク特徴量の統合により，ベースラインモデルと比較してAUCが0.62から0.66へ向上した． + 本手法は，PCIにおけるより正確な定量リスク評価の実現に寄与する可能性を示している． +\end{abstract} + +\begin{keywords} + Intravascular Ultrasound, Attenuated Plaque, CNN-LSTM +\end{keywords} + +\section{Introduction} +% 0.5〜0.75ページ + +% TODO: [図メモ] ここにIVUS画像の例を挿入予定． +% 構成案：(a) 減衰プラークが明確なフレーム，(b) 極座標変換後の画像 + +経皮的冠動脈形成術（PCI）は冠動脈疾患の標準的な治療法であるが， +プラークの流出によりNo-reflow現象やSlow flow現象といった合併症が術後に約10〜20\%の割合で発症することが報告されている\citep{babapoor2022}． +これらの合併症は一度発症すると血管拡張薬などによる事後治療が必要となるが， +急激な血圧低下や不整脈といった重篤な副作用のリスクを伴う． +そのため，発症後に対処するのではなく，合併症を未然に防ぐ予防的戦略の確立が重要である． + +現在の臨床現場では，Distal Protection Deviceが高リスク病変への有効な予防手段であるが，デバイス準備の負担から全症例への適用は現実的でなく，術前の正確なリスク層別化が求められる． +しかし，血管内超音波（IVUS）読影には専門知識を要するうえ，血管径やプラーク組成など複数の因子が複合的に絡み合うため，目視のみによる高精度な予測は困難であり，IVUS画像から定量的にリスクを評価する手法が必要とされている． + +我々の予備検討では，IVUS画像をResNet-34-long short-term memory（LSTM）モデルに入力するアプローチでarea under the curve（AUC）---閾値に依存しない指標---が0.62にとどまり，予測性能の向上が課題として残った． +IVUS上の減衰プラークとNo-reflowの有意な関連が報告されているが\citep{wu2011}，そのスコアリングは手動であり，これらの特徴を自動的に統合した予測手法が求められる． +本研究では，医師によるアノテーションから算出した減衰角度統計量をconvolutional neural network--LSTM（CNN-LSTM）に明示的に統合する手法を提案する． + + +\section{Methods} + +\subsection{Dataset} +本研究で使用するIVUSデータは，千葉大学医学部附属病院で取得されたIVUS動画から， +プラーク病変部の中心フレームを基準に1\,mmごとのスライスで前後6フレーム，計13フレームを切り出したものである． +症例数は312症例であり，そのうちNo-reflow症例が39例，Slow flow症例が35例，合併症を発症しなかった症例が238例である．No-reflow症例とSlow flow症例を合わせて陽性クラスとした． +各フレームには，医師による減衰角度方向のアノテーションが付与されている． +評価にはすべての実験において5分割交差検証を用い，各フォールドの平均値を結果として報告する． + +\subsection{Proposed Method} + +まず，各IVUSフレームに極座標変換を適用する． +変換は画像中心を原点とし，Bicubic補間により縦224 $\times$ 横254ピクセルの画像に変換する． +続いて，変換後の画像左端30ピクセルに存在するIVUSカテーテル領域を除去し， +最終的に224 $\times$ 224ピクセルの画像として入力に使用する． +この前処理により，非関心領域が除去されるとともに， +円形の血管断面が矩形領域に展開され，空間特徴の抽出が効率化される． + +次に，減衰角度のアノテーションから13フレームにわたる平均・最大値・最小値・標準偏差の4統計量を算出し，減衰プラーク特徴量とする． + +モデルはResNet-34とLSTMで構成され，LSTMの出力に減衰プラーク特徴量を結合してFC層から合併症発症確率を出力する． + +また，プラーク領域のRegion of Interest（ROI）マスクをIVUS画像にチャネル方向で結合することの寄与を検証するため，ROIマスクを追加入力とした比較手法についても評価する．各フレームのマスクは，プラーク領域のアノテーションがあるフレームには医師のアノテーションを使用し，減衰が激しくアノテーションが困難なフレームには，本データセットのアノテーション付きフレームでファインチューニングした\citet{nishi2021}のセグメンテーションモデルを用いて自動生成した． + + +\section{Results} + +\citet{wu2011}のスコアリング手法（AUC: 0.65）および同一データセットのIVUS画像から医師が行った予測（AUC: 0.56）との比較を含むアブレーションスタディの結果を\tableref{tab:results}に示す． +提案手法（AUC: 0.66）は\citet{wu2011}を上回り，CNNがIVUS画像から抽出する特徴量が減衰角度に加えて予測に有効な情報を含むことが示された． +全AIモデルはAUCにおいて医師の予測も上回った． +極座標変換のみではAUCはベースライン（0.62）と同等であったが，減衰プラーク特徴量のみの統合ではAUCが0.63に向上した． +両手法を統合した提案手法ではAUCが0.66と最高値を達成し，ベースライン（$\pm 0.12$）より分散が小さく（$\pm 0.07$），より安定した予測性能を示した． +一方，提案手法にROIマスクを追加した比較手法ではAUCが0.63にとどまり，ROIマスクの追加が必ずしも予測性能の向上に寄与しないことが示された． + +\begin{table}[htbp] + \floatconts + {tab:results}% + {\caption{アブレーションスタディの結果}}% + {\resizebox{\linewidth}{!}{\small\begin{tabular}{lccccccc} + \hline + 手法 & 極座標変換 & \makecell{減衰プラーク \\特徴量} & 正解率 & 適合率 & 再現率 & F1 & AUC \\ + \hline + \citet{wu2011} & --- & --- & $0.58 \pm 0.16$ & $0.31 \pm 0.06$ & $0.54 \pm 0.30$ & $0.36 \pm 0.10$ & $0.65 \pm 0.08$ \\ + 医師による予測 & --- & --- & $0.56 \pm 0.07$ & $0.32 \pm 0.05$ & $\mathbf{0.72} \pm 0.08$ & $\mathbf{0.44} \pm 0.06$ & $0.65 \pm 0.09$ \\ + ResNet-34-LSTM & $\times$ & $\times$ & $0.54 \pm 0.12$ & $0.31 \pm 0.08$ & $0.71 \pm 0.18$ & $0.43 \pm 0.09$ & $0.62 \pm 0.12$ \\ + ResNet-34-LSTM & $\checkmark$ & $\times$ & $0.64 \pm 0.13$ & $\mathbf{0.36} \pm 0.14$ & $0.44 \pm 0.09$ & $0.38 \pm 0.11$ & $0.62 \pm 0.12$ \\ + ResNet-34-LSTM & $\times$ & $\checkmark$ & $0.56 \pm 0.11$ & $0.31 \pm 0.07$ & $0.64 \pm 0.20$ & $0.40 \pm 0.05$ & $0.63 \pm 0.09$ \\ + ResNet-34-LSTM（提案） & $\checkmark$ & $\checkmark$ & $0.60 \pm 0.11$ & $0.35 \pm 0.10$ & $0.62 \pm 0.11$ & $0.43 \pm 0.05$ & $\mathbf{0.66} \pm 0.07$ \\ + ResNet-34-LSTM（+ROIマスク） & $\checkmark$ & $\checkmark$ & $\mathbf{0.65} \pm 0.08$ & $0.35 \pm 0.08$ & $0.53 \pm 0.22$ & $0.41 \pm 0.09$ & $0.63 \pm 0.11$ \\ + \hline + \end{tabular}}} +\end{table} + + +\section{Conclusion} +% 0.25ページ +本研究では，IVUS画像を用いたPCI合併症予測において，極座標変換と減衰プラーク特徴量を統合したCNN-LSTMパイプラインを提案した． +アブレーションスタディの結果，両手法の統合によりAUCがベースライン（0.62）から0.66へ向上し，医師による予測（AUC: 0.56）を上回る性能を達成した． +これらの結果は，IVUS画像における幾何学変換と臨床知見の明示的な統合が，深層学習モデルによる合併症の予測性能の向上に有効であることを示している． +一方，ROIマスクの追加が予測性能の向上に寄与しなかった要因として，プラーク領域のセグメンテーション精度の不足が考えられる． +今後は，石灰化プラークなど他の病変特徴や臨床情報の統合を通じた予測性能のさらなる向上が課題として挙げられる． + + +% Acknowledgments---Will not appear in anonymized version +\midlacknowledgments{本研究は千葉大学大学院医学研究倫理審査委員会の承認を得た （承認番号 3270号）．} + +\bibliography{references} + +\end{document} diff --git "a/Reference/\345\215\222\346\245\255\350\253\226\346\226\207_\351\207\221\345\255\220_\346\234\200\347\265\202.pdf" "b/Reference/\345\215\222\346\245\255\350\253\226\346\226\207_\351\207\221\345\255\220_\346\234\200\347\265\202.pdf" new file mode 100644 index 0000000..027b4a4 --- /dev/null +++ "b/Reference/\345\215\222\346\245\255\350\253\226\346\226\207_\351\207\221\345\255\220_\346\234\200\347\265\202.pdf" Binary files differ