Accurate diagnosis of idiopathic interstitial pneumonia (IIP) is crucial as each type of IIP has a different treatment and prognosis. However, accurate diagnosis can be difficult even for specialists. In this study, we developed and tested a deep learning method for segmentation and classification of IIP using computed tomography (CT).We used 37 CT scans of patients diagnosed with different IIP types, grouped as either idiopathic pulmonary fibrosis (IPF) or non-IPF. Typical patterns of IPF in CT images are called usual interstitial pneumonia (UIP). In order to segment and classify at the same time, we labeled images with specific pixel values for cases of UIP pattern and non-UIP pattern. CT images and labeled images were used to train a 3

-. The output images of 3 - were used for segmentation of IIP patterns. Classification of UIP versus non-UIP was based on the pixel ratios in the output images. The results were evaluated using leave-one-out cross validation. The Dice index, which indicates the result of segmentation, was 0.774, and the overall classification accuracy was 0.757. These results show that our method may be highly accurate for the segmentation and classification of IIP using CT images.

抄録全体を表示

Purpose: Automated analysis of skeletal muscle in whole-body computed tomography (CT) images uses bone information, but bone segmentation including the epiphysis is not achieved. The purpose of this research was the semantic segmentation of eight regions of upper and lower limb bones including the epiphysis in whole-body CT images. Our targets were left and right upper arms, forearms, thighs, and lower legs. Method: We connected two 3D U-Nets in cascade for segmentation of eight upper and lower limb bones in whole-body CT images. The first 3

- was used for skeleton segmentation in whole-body CT images, and the second 3 - was used for eight upper and lower limb bones’ segmentation in skeleton segmentation results. Thirty cases of whole-body CT images were used in the experiment, and the segmentation results were evaluated using Dice coefficient with 3-fold cross-validation. Result: The mean Dice coefficient was 93% in the left and right upper arms, 89% in the left and right forearms, 95% in the left and right thighs, and 94% in the left and right lower legs. Conclusion: Although the accuracy of the segmentation results of relatively small bones remains a challenge, the semantic segmentation of eight regions of upper and lower limb bones including the epiphysis in whole-body CT images has been achieved.

抄録全体を表示

In recent years, digital signage has become popular as a means of information dissemination to the general public. However, unlike advertisements displayed on PCs or smartphones, it is impossible to directly acquire information displayed on such signages even if the content is interesting. Mizushima et al. proposed a video watermarking method that is robust against re-shooting so that the watermark can be extracted from watermarked videos displayed on digital signage. Conventional methods have the problem of limited information capacity. In recent years, watermarking methods based on deep learning have attracted attention for embedding large watermarks. In this paper, we implemented a video electronic watermark based on 3

-, which makes it possible to embed larger watermarks than existing methods. In addition, the proposed method was able to extract the watermark from the re-shot video, and the shortest average processing time is 1.85 seconds to extract the correct watermark.

抄録全体を表示

ご遺体の死因究明に画像診断を用いる死亡時画像診断(Autopsy imaging: Ai)は，日本のCT装置の普及率の高さも相まって多くの施設で実施されてきている．現在，生体のCT画像に対する臓器抽出に関する報告は多数あるが，Aiに対する報告は少ない．本研究は，特に死後変化が大きいとされる肺領域を対象とし，生体の領域抽出法がAiに適用可能か確認するとともに，Aiにおける肺野領域を緻密に抽出することを目的とする．肺領域の抽出には，3

-を用いた．学習データは，[A]生体のCT画像55症例と，[B]生体のCT画像41症例に死後CT画像14症例を追加した2パターンを用意した．また，抽出精度はDICE係数を用いて評価した．DICE係数とは，0から1までの値をとり，抽出精度が高いほど高い値を示す．死後変化の小さいCT画像と死後変化の大きいCT画像に対して肺領域を抽出したところ，死後変化の小さい方では[A]0.79,[B]0.90,死後変化の大きい方では[A]0.69,[B]0.84であった．生体のCT画像のみを学習した[A]では，特に死後変化の大きなCT画像に対して精度よく抽出することが出来なかった．一方，変化の小さいCT画像，変化の大きいCT画像どちらに対しても[B]の方が精度よく抽出されていることが示され，学習データに死後CTを増やすことで死後変化に対応することが確認できた．

抄録全体を表示

　Pre-processing is essential for various quantitative diffusion analyses. In particular, noise reduction, Gibbs artifact removal, and distortion correction using FSL (TOPUP & EDDY) are performed. However, these pre-processing calculations take a long time. We aimed to simplify the distortion correction technique. A learning model for distortion correction was constructed by 3

- using image data before and after FSL distortion correction. Compared to Pre FSL DWI, predicted DWI improved the SSIM/PSNR and exhibited a significant difference. As a result, the learning model of FSL using 3-- was able to correct the distortions in the high b-value DWI.

抄録全体を表示

金型加工において加工効率および加工精度に大きな影響を与える工程設計の自動化に向けて，人工知能の一種である機械学習を援用した工程設計支援システムを開発している．先行研究では，ボクセルモデルを用いて仕上げ加工の工具経路パターンを推定していたが，本研究ではボクセルモデルの特徴を生かして金型の荒加工における工程情報を推定するためのネットワークを構築し，その基本的な性能を調査して有効性を検証した．

抄録全体を表示

Background: The aim of this study was to build an auto-segmented artificial intelligence model of the atria and epicardial adipose tissue (EAT) on computed tomography (CT) images, and examine the prognostic significance of auto-quantified left atrium (LA) and EAT volumes for AF.

Methods and Results: This retrospective study included 334 patients with AF who were referred for catheter ablation (CA) between 2015 and 2017. Atria and EAT volumes were auto-quantified using a pre-trained 3-dimensional (3

) - model from pre-ablation CT images. After adjusting for factors associated with AF, Cox regression analysis was used to examine predictors of AF recurrence. The mean (±SD) age of patients was 56±11 years; 251 (75%) were men, and 79 (24%) had non-paroxysmal AF. Over 2 years of follow-up, 139 (42%) patients experienced recurrence. Diabetes, non-paroxysmal AF, non-pulmonary vein triggers, mitral line ablation, and larger LA, right atrium, and EAT volume indices were linked to increased hazards of AF recurrence. After multivariate adjustment, non-paroxysmal AF (hazard ratio [HR] 0.6; 95% confidence interval [CI] 0.4–0.8; P=0.003) and larger LA-EAT volume index (HR 1.1; 95% CI 1.0–1.2; P=0.009) remained independent predictors of AF recurrence.

Conclusions: LA-EAT volume measured using the auto-quantified 3

- model is feasible for predicting AF recurrence after CA, regardless of AF type.

抄録全体を表示

メラノーマは、悪性度の高い皮膚がんである。初期病変は良性病変との区別が非常に難しい。そのため、精度が高く定量的な診断手法に対するニーズは強い。これまでにも、メラノーマの診断支援を目的としたシステムが多く開発されている。また、これまでの知見から、皮膚色素性病変画像から病変部を抽出すると、診断性能が向上することが分かっている。本研究では、多波長分光装置の一種であるHyperspectral Imager(HSI)によって撮影された皮膚色素病変データを対象とする。HSIによって計測されるHyperspectral Data(HSD)は高さ×幅×波長の3次元データ構造を持つため、一般的なセグメンテーションモデルであるu-netをそのまま使用することはできない。そこで、u-netを3次元に拡張した3

-を採用し、病変部の抽出を試みる。国内と北欧で撮影された234例のメラノーマと285例の非メラノーマのHSDと、皮膚科専門医による病変抽出画像を対象とする。HSDの取得は倫理審査委員会の承認を得たプロトコールに従って実施された。3 -による抽出結果を解析した結果、jaccard係数0.613、F1スコア0.714が得られた。今後は、HSDに正規化や標準化などの前処理を施すことや、Googleによって開発されたセグメンテーションモデルであるDeeplab v3+の採用や、2次元画像を入力としたu-netとの比較などを行う予定である。

抄録全体を表示

本稿では，スーパーコンピューターを活用したディープラーニング（DL）を用いたコンピューター支援検出（CAD）ソフトウェアの開発について紹介する．スーパーコンピューター上に構築した非同期並列実行型ベイズ的最適化に基づくDL 学習環境を紹介するとともに，3

- を用いた胸部CT 画像の肺結節検出におけるハイパーパラメーター探索の例を示した．構築した学習環境により，ハイパーパラメーター探索を伴うDL の学習の期間短縮が実現できた．

抄録全体を表示

高精度かつ高速な医療画像のセグメンテーションは，多くの医療現場において重要な課題である．現在ではその手法の一つとして，エネルギー最小化問題に基づくグラフカットが利用されている．しかし，グラフカットでは，隣接するピクセル値が類似している場合，完全かつ自動的にセグメンテーションを行うことは困難である．この問題には多くの対策があるが，そのほとんどは実行速度という点で適していない．それに対し，深層学習による手法は，複雑な特徴を獲得することができるため，自動セグメンテーションが可能である．本研究では，Residual Unitによって拡張した3

-と，セグメンテーション結果を修正する3DCNNを組み込んだモデルを提案する．

抄録全体を表示

Dies and molds are necessary elements in the manufacturing of current industrial products. There is increasing pressure to machine high quality complicated surfaces at low cost. The standardization of process planning is said to be a key to improving the efficiency of machining operations in practice. Thus, computer aided process planning (CAPP) systems are urgently needed to reduce the time and effort of preparing machining operations. However, it is difficult to generalize process planning that continues to depend on skillful experts and requires long preparation time for die and mold machining. On the other hand, to overcome issues that are difficult to generalize, it is well known that machine learning has the capability to estimate valid values according to past case data. Therefore, this study aims to develop a CAPP system that can determine machining process information for complicated surfaces of die and mold based on pattern recognition and deep learning, a kind of machine learning. A network architecture called 3

- is adapted to effectively analyze whole images by producing segmented regions. Using a voxel model representing targeted shape, it becomes easier to deal with the complicated surfaces of die and mold generally and three-dimensionally, as skilled experts pay attention to whole geometrical features. Cutting tool type and tool path pattern are treated as machining process information determined in a CAPP system. The results of case studies confirm that the developed CAPP system is effective in determining the machining process information even for complicated surfaces according to the implicit machining know-how.

抄録全体を表示

【目的】本研究では、CT値低下が著しく骨として認識しにくい溶骨性骨転移部を見落とすことなく高精度に全身FDG-PET/CT（以降PET/CT）像上の骨領域を抽出する方法として、PETとCTの断面像の同時解析と骨の空間連続性を考慮可能な2.5次元 (2.5D) 処理を用いた骨領域抽出法を提案する。

【手法】提案法は、2.4mm空間等方化、画素値正規化 (SUV：0~5、CT値：WL＝200、WW＝2000)、顎下以上除去を施したPET/CT像を対象とする。まず、PET/CT 像のAxial、Coronal、Sagittal断面像それぞれを処理する3つの2ch入力2

- を用い、断面像上の骨抽出結果を得る。次に、得られた3断面の骨抽出結果を平均値統合し3D骨領域を得る。最後に得られた領域から小成分を除去する。

【結果】全身PET/CT像40症例を用いた5-fold交差検証実験の結果、提案法による抽出骨領域の制度し尿Dice Indexは平均0.960、標準偏差0.006、溶骨性転移領域の抽出感度は平均0.968、標準偏差0.047であった。この抽出精度は、Axial断面のみの2

-処理による骨抽出結果に比べて統計的有意に高い値であった。

【結語】臨床画像を用いた検証から、提案するU-Net 2.5D処理によるPET/CT像上骨領域抽出法の有用性を確認した。

抄録全体を表示

Purpose: To analyze subcortical brain volume more reliably, we propose a deep learning segmentation method of subcortical brain based on magnetic resonance imaging (MRI) having high generalization performance, accuracy, and robustness.

Methods: First, local images of three-dimensional (3D) bounding boxes were extracted for seven subcortical structures (thalamus, putamen, caudate, pallidum, hippocampus, amygdala, and accumbens) from a whole brain MR image as inputs to the neural network. Second, dilated convolution layers, which input information of variable scope, were introduced to the blocks that make up the neural network. These blocks were connected in parallel to simultaneously process global and local information obtained by the dilated convolution layers. To evaluate generalization performance, different datasets were used for training and testing sessions (cross-dataset evaluation) because subcortical brain segmentation in clinical analysis is assumed to be applied to unknown datasets.

Results: The proposed method showed better generalization performance that can obtain stable accuracy for all structures, whereas the state-of-the-art deep learning method obtained extremely low accuracy for some structures. The proposed method performed segmentation for all samples without failing with significantly higher accuracy (P < 0.005) than conventional methods such as 3

-, FreeSurfer, and Functional Magnetic Resonance Imaging of the Brain’s (FMRIB’s) Integrated Registration and Segmentation Tool in the FMRIB Software Library (FSL-FIRST). Moreover, when applying this proposed method to larger datasets, segmentation was robustly performed for all samples without producing segmentation results on the areas that were apparently different from anatomically relevant areas. On the other hand, FSL-FIRST produced segmentation results on the area that were apparently and largely different from the anatomically relevant area for about one-third to one-fourth of the datasets.

Conclusion: The cross-dataset evaluation showed that the proposed method is superior to existing methods in terms of generalization performance, accuracy, and robustness.

抄録全体を表示

多くの工業製品に用いられる金型は複雑形状化が進み，低コストで高品位な意匠面を加工することが求められている．しかし，加工精度やコストに大きく影響を与える工程設計は未だに熟練技能者に依存しており，一般化も困難である．そこで本研究では，ボクセルモデルとDeep Learningによる金型加工の工程設計支援システムの開発の第一歩として，意匠面の工具経路パターンの推定に取り組み，ケーススタディによって有効性を検証した．

抄録全体を表示

Lung cancer is the most common type of cancer and is the leading cause of cancer-related deaths in Japan. Regarding lung cancer diagnosis, a pivotal aspect in lung cancer treatment is tumor detection and selection of appropriate cancer treatment. Computer Tomography (CT) images are usually used to detect tumors. After tumor detection, identifying mutations in the Epidermal Growth Factor Receptor (EGFR) gene is essential for cancer treatment selection. The EGFR gene is a key factor associated with cancer cell proliferation its mutation appears both inside and around the tumors. This paper proposes a lung cancer diagnostic system designed to streamline the process from tumor detection to EGFR gene mutation identification. The proposed system consists of three modules: an input interface module, an automated lung tumor segmentation module, and an EGFR mutation prediction module. The system is characterized in that all modules are consistently based on image processing. Consequently, the proposed system enables users to acquire the diagnosis results of tumor detection as well as EGFR gene mutation prediction by just providing the input CT image data and the patients' clinical features. Our experimental results confirm that the system achieves performance levels comparable to existing research, both in terms of lung tumor segmentation accuracy and the precision of EGFR mutation predictions.

抄録全体を表示

One of the most common tasks in medical imaging is semantic segmentation. Achieving this segmentation automatically has been an active area of research, but the task has been proven very challenging due to the large variation of anatomy across different patients. However, recent advances in deep learning have made it possible to significantly improve the performance of image recognition and semantic segmentation methods in the field of computer vision. Due to the data driven approaches of hierarchical feature learning in deep learning frameworks, these advances can be translated to medical images without much difficulty. Several variations of deep convolutional neural networks have been successfully applied to medical images. Especially fully convolutional architectures have been proven efficient for segmentation of 3D medical images. In this article, we describe how to build a 3D fully convolutional network (FCN) that can process 3D images in order to produce automatic semantic segmentations. The model is trained and evaluated on a clinical computed tomography (CT) dataset and shows stateof-the-art performance in multi-organ segmentation.

抄録全体を表示

In this paper, we present a deep learning-based method for kidney and tumor region segmentation using 3D CT data from multiple sources. We conduct experiments by training with mixed datasets and fine-tuning transfer learning. Throughout these experiments, data augmentation is applied by blending arterial phase CT images and portal vein phase CT images. Our findings reveal a trend: higher accuracy in predicting kidney labels is achieved when fine-tuning transfer learning is applied, while higher accuracy in predicting tumor labels is attained when training with a mixed dataset. This suggests the effectiveness of fine-tuning when the variations in the datasets are relatively small, as seen in the case of kidneys. Conversely, training with mixed datasets proves effective when the variation in prediction targets is relatively large, such as with tumors. It is also confirmed that integrating the results by different training policies improves the overall segmentation results.

抄録全体を表示

Convolutional neural networks (CNNs) have become popular in medical image segmentation. The widely used deep CNNs are customized to extract multiple representative features for two-dimensional (2D) data, generally called 2D networks. However, 2D networks are inefficient in extracting three-dimensional (3D) spatial features from volumetric images. Although most 2D segmentation networks can be extended to 3D networks, the naively extended 3D methods are resource-intensive. In this paper, we propose an efficient and accurate network for fully automatic 3D segmentation. Specifically, we designed a 3D multiple-contextual extractor to capture rich global contextual dependencies from different feature levels. Then we leveraged an ROI-estimation strategy to crop the ROI bounding box. Meanwhile, we used a 3D ROI-attention module to improve the accuracy of in-region segmentation in the decoder path. Moreover, we used a hybrid Dice loss function to address the issues of class imbalance and blurry contour in medical images. By incorporating the above strategies, we realized a practical end-to-end 3D medical image segmentation with high efficiency and accuracy. To validate the 3D segmentation performance of our proposed method, we conducted extensive experiments on two datasets and demonstrated favorable results over the state-of-the-art methods.

抄録全体を表示

【背景・目的】我が国では3D MRI画像から内側側頭部脳領域の萎縮度を評価するVSRADが脳ドック等で広く使われてきた。2020年に発売されたSYNAPSE VINCENTの脳区域解析アプリは脳と脳脊髄液をAIで26区域に自動分割できる。2022年より運用が始まったクラウド型AI開発支援サービスSYNAPSE Creative Space (CS)は特定関心領域を抽出し、病的所見の有無を自動判定するAIを医療者自らが作成できる。これらを活用した医療・脳ドックの未来を紹介する。【方法】健常成人133人に3T MRIで3D T1-MPRAGE画像を撮影。脳区域解析アプリで、前頭葉、頭頂葉、側頭葉、後頭葉、海馬、小脳、脳幹等21領域と、各脳室、くも膜下腔の5区域に自動分割、体積割合を算出し、局所脳と髄液の加齢性変化を検証。SYNAPSE CSで特発性正常圧水頭症(iNPH) に特徴的な画像所見DESHを自動判別すべく、頭蓋内CSF、脳室とくも膜下腔2領域の領域分割と脳室拡大(VD)、高位円蓋部・正中くも膜下腔狭小化(THC)、シルビウス裂開大(SFD)の自動判定を推論。【結果】20歳以降加齢とともに大脳皮質灰白質は直線的に減少し、大脳辺縁系等の皮質下灰白質は維持され、白質は40代までわずかに増加後、急速に減少。一方、くも膜下腔は直線的に増加し、脳室は60代以降に増加。SYNAPSE CS によるiNPHのDESH、VD、THC、SFDの定量的自動判定は十分な精度であった。【結語】従来、「脳委縮」や「DESH」は医師の主観で評価されてきたが、今後はAIを活用した各脳区域の体積割合による脳委縮やDESHの自動判定など、数値で客観的に評価される。

抄録全体を表示