An innovative neural network framework to classify blood vessels and tubules based on Haralick features evaluated in histological images of kidney biopsy

详细信息    查看全文

• 作者：Vitoantonio Bevilacqua^a ; ^{vitoantonio.bevilacqua@poliba.it}Author Vitae ; Nicola Pietroleonardo^a ; ^bAuthor Vitae ; Vito Triggiani^a ; ^bAuthor Vitae ; Antonio Brunetti^aAuthor Vitae ; Anna Maria Di Palma^bAuthor Vitae ; Michele Rossini^bAuthor Vitae ; Loreto Gesualdo^bAuthor Vitae
• 关键词：Computer Aided Diagnosis ; Supervised artificial neural network ; Image processing and segmentation ; Vessels detection ; Histological Image ; Haralick features
• 刊名：Neurocomputing
• 出版年：2017
• 出版时间：8 March 2017
• 年：2017
• 卷：228
• 期：Complete
• 页码：143-153
• 全文大小：1214 K
• 卷排序：228

文摘

Introduction and objective: Computer Aided Diagnosis (CAD) systems based on Medical Imaging could support physicians in several fields and recently are also applied in histopathology. The aim of this work is to discuss in detail the design and testing of a CAD system for segmentation and discrimination of blood vessels versus tubules from biopsies in the kidney tissue through the elaboration of histological images.Materials and methods: Materials consist in 10 Kidney Biopsy Slides (KBS) with Periodic Acid Schiff (PAS) staining. The Regions Of Interest (ROI) identified by expert are in total 221:71 vessels and 150 tubules. KBS preparation and digital acquisition have been conducted by expert technicians at the Department of Emergency and Organ Transplantation (DETO) of the University of Bari Aldo Moro (Italy). Each slice is a Red Green Blue (RGB) format image with a resolution of 0.50 µm/pixel. Starting from KBS images, several techniques were tested for ROI's segmentation and classification. In particular, we formerly describe the innovative preliminary step to segment regions of interest, the procedure to extract significant features from them and finally discuss the supervised Artificial Neural Networks (ANNs) architecture based on error back propagation training algorithm. All the training sets were builts by using vessels and non vessels (tubules) ROI samples, whose dimensions were correlated to the vessels to be detected.Results: The performance of the best ANN architecture, trained by using a training set of 35 vessels among the 71 available vessels in dataset, were evaluated in terms of False Positives (FPs) and False Negatives (FNs). On an initial reduced dataset, it reveals good performance and robustness in terms of FPs reduction.Conclusion: Tests determined that the supervised ANN approach is consistent and reveals good performance, after a training phase based on vessels and non-vessels (tubules) samples. Moreover, our method could be improved by using a larger dataset diagnosed by expert nephropathologists.