Examination of various classification strategies in classification of lung nodule characteristics


Creative Commons License

KAYA A., KEÇELİ A. S., CAN A. B.

JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY, cilt.34, sa.2, ss.710-725, 2019 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 34 Sayı: 2
  • Basım Tarihi: 2019
  • Doi Numarası: 10.17341/gazimmfd.416530
  • Dergi Adı: JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, TR DİZİN (ULAKBİM)
  • Sayfa Sayıları: ss.710-725
  • Anahtar Kelimeler: Nodule characteristics, pulmonary nodules, transfer learning, deep features, image processing, PULMONARY NODULES, DATABASE
  • Hacettepe Üniversitesi Adresli: Evet

Özet

Nodule characteristics used in the evaluation of lung nodules are generally subjective assessments of the expert opinions. Among these characteristics, the most known and used one for prediction is the degree of malignancy. In the classification studies in the literature, deep features are used besides the traditional features extracted from the nodule appearance and morphological structure. In this study, traditional features, deep features, and the combinations both used in predicting the nodule characteristics. Four classification algorithms with different structures are evaluated for predicting nodule characteristics. Reference data sets of nodule characteristics were generated by means of majority voting from subjective assessments of radiologists. These generated data sets generally have large unbalanced class distributions. Data balancing procedure has been applied to examine the effect of this condition on classification results. With the combinations of these methods, effects of different classification models on the classification accuracy, sensitivity and specificity are examined. The results of the experiments shown that the classification strategy needs to be specifically determined starting from the used features to the classification algorithm according to the performance criterion to be achieved.