基於卷積神經網路與遷移學習的兩階段結核菌培養檢測

Abstract

肺結核目前是全球第十大死因，嚴重威脅了全人類，尤其是發展中國家的公共衛生。為了能夠及早診斷治療肺結核，自動化電腦輔助診斷系統變得十分重要。本文提出了一個基於卷積式神經網路及遷移學習的電腦輔助診斷系統，用以分類肺結核檢驗之黃金標準—痰液培養檢測—的培養結果。結合遷移學習以及本文提出的資料切割方法，我們克服了小且不平衡的醫療資料集為卷積式神經網路訓練過程所帶來的挑戰。另外，由於「非陰性類別」的表現在醫療檢測領域上十分重要，我們設計了一個兩階段分類方法來改善「非陰性類別」的分類結果。我們以來自台灣桃園醫院的16,503張真實痰液培養影像來驗證本研究的貢獻。實驗結果顯示，本文提出的系統在非陰性類別達到了98%的召回率以及99%的精準率。 （因本研究正商談技轉中，故以CONFIDENTIAL閉門口試進行。）

Tuberculosis (TB) is the tenth cause of death in the world. It has seriously threatened human health, especially in developing countries. A computer-aided diagnosis (CADx) system is necessary to accelerate the TB diagnosis for early treatment. In this thesis, we propose a CADx system using convolutional neural network and transfer learning to classify the result of “TB culture test” - the gold standard for TB diagnostic test. We apply transfer learning and propose a data splitting method to resolve the challenge of the small and imbalanced dataset. In addition, as the performance of non-negative class is crucial in this application, we introduce a two-stage classification method (TSCM) to boost the results. Experiments use a real clinical dataset of TB culture test (16,503 images from Tao-Yuan General Hospital, Taiwan) to verify our work. Our proposed method achieves 98% recall and 99% precision on non-negative class. (This thesis is CONFIDENTIAL with a defence behind closed doors as its technical transfer is still under discussion.)