Signal and Data Processing

fa تشخیص چندکلاسۀ بیماری‌های ریوی از روی تصاویر بهبود‌یافتۀ رادیوگرافی قفسۀ سینه مبتنی‌بر یادگیری تجمیعی Multi-class Lung Disease Detection from Enhanced CXR Images Using Ensemble Learning مقالات گروه علائم حیاتی ( مرتبط با مهندسی پزشکی) Paper پژوهشي Research <div style="text-align: justify;">در این مقاله، یک روش ترکیبی طبقه‌بندی چندکلاسه مبتنی‌بر الگوریتم‌های پیشرفته یادگیری عمیق ارائه شده‌است. پایگاه داده COVID-19، که شامل مجموعه‌ای از تصاویر رادیوگرافی قفسه سینۀ<a href="#_ftn1" name="_ftnref1" title="">[1]</a> عمومی است، برای بررسی تأثیر روش‌‌های تصویربرداری ساده، بهبودیافته و بخش‌بندی‌شده مورد استفاده قرار گرفته است؛ علاوه‌براین، مزایای نهفته ترکیب تصاویر بخش‌بندی‌شده -بهبودیافته و بهبودیافته- بخش‌بندی‌شده نیز بررسی شده‌است؛ همچنین، به‌طور خاص تأثیر ترتیب اجرای مراحل بهبود و بخش‌بندی نیز ارزیابی شده‌است. بخش‌بندی ریه با استفاده از مدل U-Net انجام شده‌است؛ درحالی‌که بهبود تصاویر خام و بخش‌بندی‌شدۀ رادیوگرافی قفسه سینه با استفاده از یک شبکه عصبی پیشرفته صورت گرفته است. در مرحله پیش‌پردازش، تصاویر رادیوگرافی قفسه سینه تغییر اندازه داده شده و نرمال‌سازی می‌شوند؛ سپس به هفت مدل یادگیری انتقالی شامل VGG16، VGG19، ResNet50V2، InceptionV3، MobileNet، InceptionResNetV2 و DenseNet201 ورودی داده می‌شوند. قدرت این مدل‌ها از طریق یک رویکرد گروهی ترکیب شده و چهارچوب بهبودیافته پیشنهادی عملکرد برتری را با دقت 97،03 درصد نشان داده و از پژوهش‌های اخیر موجود پیشی گرفته است.</div> <div>  <hr align="left" size="1" width="33%" > <div dir="ltr" id="ftn1"><a href="#_ftnref1" name="_ftn1" title="">[1]</a> Chest X-Ray (CXR)</div> </div> <div class="WordSection1">Every year, millions of people around the world suffer from lung diseases like COVID-19, pneumonia, and lung opacity. These illnesses are among the leading causes of death, and early diagnosis is critical for effective treatment. Chest X-rays (CXR) are commonly used because they are cheap and widely available, but they often contain noise and unclear features, which can lead to mistakes even by experienced radiologists. In this study, we wanted to see whether we could improve the accuracy of automatic lung disease detection by enhancing image quality and segmenting (isolating) the lungs. More importantly, we were curious about the order of these steps: does it matter if we enhance first and then segment, or segment first and then enhance? To answer this, we designed and compared five different image processing pipelines. Methodology: We used a public dataset called the COVID-19 Radiography Database, which contains CXR images of four categories: COVID-19, viral pneumonia, lung opacity, and normal. To avoid bias, we balanced the dataset so that each class had exactly 1,345 images. For lung segmentation, we trained a U-Net model on another dataset of 566 X-ray images with their corresponding lung masks. This model performed very well, achieving a Dice coefficient of 0.96 and an IoU of 0.93. For image enhancement, we used a special type of neural network called a Residual Dense Neural Network (RDNN), which sharpens edges and reduces noise. We then created five different pipelines: <ol> <li class="CxSpMiddle" style="text-align:justify">Plain images (no preprocessing)</li> <li class="CxSpMiddle" style="text-align:justify">Enhanced images only</li> <li class="CxSpMiddle" style="text-align:justify">Segmented images only</li> <li class="CxSpMiddle" style="text-align:justify">Enhanced first, then segmented</li> <li class="CxSpMiddle" style="text-align:justify">Segmented first, then enhanced</li> </ol> For each pipeline, we fed the images into seven well-known deep learning models (VGG16, VGG19, ResNet50V2, InceptionV3, MobileNet, InceptionResNetV2, and DenseNet201) that had been pre-trained on a large dataset called ImageNet. We used these models as feature extractors – meaning they pulled out important visual features from the X-rays. Then, instead of relying on just one model, we combined their predictions using an ensemble method called a meta-classifier, which learns to make the final decision based on the strengths of all seven models. Results: The plain images (pipeline 1) gave us an accuracy of 96.38%, which is already quite good. But pipeline 2 – the enhanced images – performed even better, reaching 97.03% accuracy, along with 97.12% precision, 97.03% recall, 97.1% F1-score, and an AUC of 98.98%. This was our best result. Interestingly, when we added segmentation to the mix, performance dropped. Pipeline 3 (segmented only) achieved only 88.84% accuracy. Even more surprising, pipeline 4 (enhanced then segmented) and pipeline 5 (segmented then enhanced) gave even lower accuracies of 79.93% and 86.25%, respectively. Looking at the confusion matrices, we noticed that the most common mistakes were confusing lung opacity with pneumonia – which makes sense because these two conditions can look very similar on X-rays even to human eyes. Conclusion: This study shows that simply enhancing CXR images with a residual dense neural network, followed by an ensemble of seven transfer learning models, can achieve state-of-the-art results (97.03% accuracy) for detecting multiple lung diseases. Surprisingly, adding lung segmentation did not help, it actually made things worse, depending on the order of operations. This finding is important because many researchers assume that segmentation always improves performance, but our results suggest otherwise for this particular task. Of course, our study has limitations: the dataset was still relatively small, and we only used X-rays (not CT scans). In the future, we plan to test our method on larger and more diverse datasets, include CT images, use data augmentation to reduce overfitting, and improve the segmentation model so it works better with enhanced images. </div> <div aria-label="Page Break" class="cke_pagebreak" contenteditable="false" data-cke-display-name="pagebreak" data-cke-pagebreak="1" style="page-break-after:always" title="Page Break"></div> بیماری‌های ریوی, تصاویر رادیوگرافی قفسه سینه, یادگیری تجمیعی, یادگیری انتقالی, پردازش تصویر. Lung diseases, chest X-ray images, ensemble learning, transfer learning, image processing. 61 82 http://jsdp.rcisp.ac.ir/browse.php?a_code=A-10-2187-2&slc_lang=fa&sid=1 Shahriar Mohammadi شهریار محمدی mohammadi@kntu.ac.ir Yes Associate Professor, Department of Industrial Engineering, KN Toosi University of Technology, Tehran, Iran دانشیار، دانشکدۀ مهندسی صنایع، دانشگاه خواجه نصیرالدین طوسی، تهران، ایران Zahra Taghavian Asiabi زهرا تقویان آسیابی z.taghavian@email.kntu.ac.ir No Master Degree, Department of Industrial Engineering, KN Toosi University of Technology, Tehran, Iran دانش‌آموختۀ کارشناسی‌ارشد، دانشکدۀ مهندسی صنایع، دانشگاه خواجه نصیرالدین طوسی، تهران، ایران Mehdi Babagoli مهدی باباگلی babagoli@umz.ac.ir No Assistant Professor, Department of Computer Engineering, University of Mazandaran, Babolsar, Iran استادیار، دانشکدۀ مهندسی کامپیوتر، دانشگاه مازندران، بابلسر، مازندران، ایران