Authors:
Hendra Bunyamin
1
;
Hapnes Toba
1
;
Meyliana
2
and
Roro Wahyudianingsih
3
Affiliations:
1
Informatics Engineering, Maranatha Christian University, Jl Prof. drg. Surya Sumantri, M.P.H. No. 65, Bandung, Indonesia
;
2
Accounting Department, Maranatha Christian University, Jl Prof. drg. Surya Sumantri, M.P.H. No. 65, Bandung, Indonesia
;
3
Faculty of Medicine, Maranatha Christian University, Jl Prof. drg. Surya Sumantri, M.P.H. No. 65, Bandung, Indonesia
Keyword(s):
Breast Cancer Histopathological Image Classification, Deep Learning, Convolutional Neural Network, Progressive Resizing, Vahadane Image Normalization.
Abstract:
Breast cancer (BC) is a lethal disease which causes the second largest number of deaths among women in the world. A diagnosis of biopsy tissue stained with hematoxylin & eosin, commonly named BC histopathological image, is a non-trivial task which requires a specialist to interpret. Recently, the advance in machine learning techniques driven by deep learning techniques and competition datasets has enabled the automation and prediction of histopathological images interpretation. Each different competition dataset has its own state-of-the-art technique; therefore, this paper explores an avenue of research by merging popular BC histopathological images research datasets and searching for the best performing models on the unified dataset. The merging process maintains similar classes among the datasets; consequently, the unified dataset has three classes and the prediction problem is cast into multi-class classification problem. We propose a combination of Vahadane preprocessing techniqu
e and training method using progressive resizing approach. Our approach demonstrates that both utilizing Vahadane image normalization and utilizing our progressive resizing technique achieve around 99% in F1 score , which is comparable among other state-of-the-art models. The unified dataset is also provided online for advancing research in histopathological images interpretation.
(More)