Computer aided detection of prostate cancer using multiwavelength photoacoustic data with convolutional neural network

• Published in: Biomedical signal processing and control 2020-07, Vol.60, p.101952, Article 101952
• Main Authors: Dhengre, Nikhil, Sinha, Saugata, Chinni, Bhargava, Dogra, Vikram, Rao, Navalgund
• Format: Article
• Language: English
• Subjects: Computer aided detection; Convolutional neural network; Multi-wavelength photoacoustic imaging; Photoacoustic imaging; Prostate cancer
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2020.101952

•Demonstration of the effectiveness of CNN learned features from raw photoacoustic data for prostate tissue characterization.•Raw photoacoustic data is generated from excised prostates of human patients.•Tissue characterization experimentation is done using classifiers with CNN learned features and raw photoacoustic features.•Results suggest that CNN based classifiers performed better than the classifiers applied with raw photoacoustic features. In a conventional computer aided diagnosis workflow, features/markers, which are extracted from regions of interest in medical images, are analysed and assigned to different classes corresponding to normal and diseased organ. A convolutional neural network (CNN) based classifier can autonomously extract discriminative features from the medical images and then perform classification using these extracted features. The aim of this study was to evaluate the performance of autonomously learned features from photoacoustic data by the convolution layer of a CNN for differentiating between different tissue pathologies. In this study, CNN based classifiers were trained with photoacoustic data, generated from freshly excised prostates of 30 human patients who went prostatectomy for biopsy confirmed prostate cancer. Three different photoacoustic datasets, acquired at 760 nm, 800 nm and 850 nm wavelengths and the combination of all three datasets, were employed to autonomously learn discriminative features by CNN and these features were utilised with different classifiers for differentiating among malignant prostate, benign prostatic hyperplasia and normal prostate tissue. The performance of these classifiers was compared with the performance of classifiers applied with raw photoacoustic data. Two out of three CNN based classifiers provided accuracy as well as sensitivity values which were approximately equal to or higher than 0.92 for malignant versus non malignant prostate tissue classification and malignant versus normal prostate tissue classification using the combined photoacoustic dataset. The preliminary results of this study show that features learned by CNN can be successfully used for efficient prostate tissue characterisation using Photoacoustic data.