Radyogenomik veri setleri için yapay zekâ tabanlı bir hassas tıp karar destek sisteminin geliştirilmesi
Küçük Resim Yok
Tarih
2025
Yazarlar
Dergi Başlığı
Dergi ISSN
Cilt Başlığı
Yayıncı
İnönü Üniversitesi
Erişim Hakkı
info:eu-repo/semantics/openAccess
Özet
Amaç: Bu çalışmanın amacı, KHDAK hastaların biyomedikal görüntüleri ve ilgili klinik veriler kullanılarak derin öğrenme yöntemine dayalı DICOM görüntülerden süper piksel segmentasyon, otsu eşikleme ve hastalığın referans konumu belirlenmiştir. Görüntülerden alınan nicel veriler klinik veriler ile entegre edildi. Biyobelirteçlerin belirlenmesi için çeşitli makine öğrenme yöntemleri uygulanarak sınıflandırma performanslarını, klinik veriler, kilinik veriler+görüntü verileri ve modele ilişkin iyileştirme oranlarını belirlemektir. Materyal ve Metot: Çalışmada kullanılan klinik veri setinde, EGFR mutasyonuna sahip 43, mutasyon olamayan 168 ve KRAS mutasyonuna sahip 38, mutasyon olmayan 173 birey olmak üzere her bir mutasyonda 211'er KHDAK bulunmaktadır. İşlenen görüntü sayısı 2.231 ve her bir görüntü VGG16 derin öğrenme modeli ile 25.088 değişken çıkarıldı. Sınıflandırma algoritmalardan, XGBoost, CatBoost, Random Forest, DVM modelleri uygulanmıştır. Bulgular: Küçük hücreli Dışı Akciğer kanseri hastaların EGFR ve KRAS mutasyon durumları biyomedikal klinik verilerde anlamlı farklılıklar bulunmuştur. Random forest yöntemi ile en önemli 50 değişken seçilerek model eğitiminde kullanılmıştır. En iyi sınıflandırma model performansının XGBoost, CatBoost yöntemlerine ait olduğu belirlenmiş ve bu modellere ilişkin sınıflandırma performansları doğruluk, dengelenmiş doğruluk, kesinlik, duyarlılık, F1-skoru, ROC-AUC metrikleri ile değerlendirilmiş ve bunlar sırasıyla, (0.965 ± 0.015), (0.954 ± 0.021), (0.953 ± 0.024), (0.994 ± 0.007), (0.973 ± 0.011) ve (0.990 ± 0.005) olarak hesaplanmıştır. Sonuç: Gerçekleştirilen çalışmada makine öğrenmesi yöntemleri sonuçlarından hareketle, görüntü verilerinde KRAS mutasyon durumunu tahmin etmede etkili olduğu XGBoost ve CatBoost modelleri daha başarılı performans gösterdiği saptanmıştır. EGFR mutasyon durumunu tahmin etmede CatBoost daha başarılı performans göstermiştir. Anahtar Kelimeler: Akciğer Kanseri, Biyomedikal, Dijital Görüntüleme, Makine Öğrenme
Aim: This study's objective was to use biomedical images of NSCLC patients and associated clinical data to use a deep learning approach to identify super pixel segmentation, herbaceous thresholding, and disease reference location from DICOM images. Clinical data and quantitative information from image were combined. Biomarker identification, classification performance, clinical data, clinical data + image data, and model improvement rates were all assessed using a variety of machine learning techniques. Materials and Methods: The study's clinical dataset included 211 NSCLCs with each mutation, including 38 persons with a KRAS mutation and 173 without, and 43 individuals with an EGFR mutation and 168 without. A total of 2,231 photos were analyzed, and 25,088 features were retrieved from each image using the VGG16 deep learning model. XGBoost, CatBoost, Random Forest, and SVM models were utilized among the classification techniques. Results: EGFR and KRAS mutation status of lung cancer patients were found to have significant differences in biomedical clinical data. The most important 50 features were selected by random forest method and used in model training. It was determined that the best classification model performance belonged to XGBoost, CatBoost methods and the classification performances of these models were evaluated with accuracy, balanced accuracy, precision, sensitivity, F1-score, ROC-AUC metrics, respectively, (0. 965 ± 0.015), (0.954 ± 0.021), (0.953 ± 0.024), (0.994 ± 0.007), (0.973 ± 0.011) and (0.990 ± 0.005), respectively. Conclusion: The results of the study's machine learning techniques indicate that both XGBoost and CatBoost models can accurately identify the presence of KRAS mutations in image data, with XGBoost and CatBoost models performing better than the others. CatBoost was better at predicting the presence of EGFR mutations. Keywords: Biomedical, Digital Imaging, Lung Cancer, Machine Learning.
Aim: This study's objective was to use biomedical images of NSCLC patients and associated clinical data to use a deep learning approach to identify super pixel segmentation, herbaceous thresholding, and disease reference location from DICOM images. Clinical data and quantitative information from image were combined. Biomarker identification, classification performance, clinical data, clinical data + image data, and model improvement rates were all assessed using a variety of machine learning techniques. Materials and Methods: The study's clinical dataset included 211 NSCLCs with each mutation, including 38 persons with a KRAS mutation and 173 without, and 43 individuals with an EGFR mutation and 168 without. A total of 2,231 photos were analyzed, and 25,088 features were retrieved from each image using the VGG16 deep learning model. XGBoost, CatBoost, Random Forest, and SVM models were utilized among the classification techniques. Results: EGFR and KRAS mutation status of lung cancer patients were found to have significant differences in biomedical clinical data. The most important 50 features were selected by random forest method and used in model training. It was determined that the best classification model performance belonged to XGBoost, CatBoost methods and the classification performances of these models were evaluated with accuracy, balanced accuracy, precision, sensitivity, F1-score, ROC-AUC metrics, respectively, (0. 965 ± 0.015), (0.954 ± 0.021), (0.953 ± 0.024), (0.994 ± 0.007), (0.973 ± 0.011) and (0.990 ± 0.005), respectively. Conclusion: The results of the study's machine learning techniques indicate that both XGBoost and CatBoost models can accurately identify the presence of KRAS mutations in image data, with XGBoost and CatBoost models performing better than the others. CatBoost was better at predicting the presence of EGFR mutations. Keywords: Biomedical, Digital Imaging, Lung Cancer, Machine Learning.
Açıklama
Anahtar Kelimeler
Biyoistatistik, Biostatistics
