A deep learning approach for breast tissue malignancy diagnosis using micro-FTIR hyperspectral imaging
Carregando...
Data
Data de publicação:
2021
Orientador
Título da Revista
ISSN da Revista
Título do Volume
É parte de
É parte de
É parte de
É parte de
ENCONTRO DE OUTONO DA SOCIEDADE BRASILEIRA DE FÍSICA, 44.
Exportar
Resumo
The breast cancer is the most incident cancer in women with an estimative of 2.1 million new cases in 2018. With the grown of deep learning techniques, several approaches in vibrational spectroscopy have been studied. In this way, this work aimed to classify breast samples as breast cancer or adenosis using a deep learning model. It was used the human breast cancer microarray BR804b (Biomax, Inc., USA), where one core of each group, cancer and adenosis, was imaged by a Cary Series 600 micro-FTIR imaging system (Agilent Technologies, USA). The system has a spatial resolution of 5.5 μm and about 100 thousand spectra were acquired for each group. The regions of interest were selected by two k-means clustering using amide I/II (1700 to 1500 cm-1) and highest paraffin intensity (1480 to 1450 cm-1) bands. Spectra were preprocessed by five steps: outlier removal using Hotelling’s T2 versus Q residuals; biofingerprint truncation; Savitzky–Golay filtering for smoothing and second derivative; Extended multiplicative signal correction (EMSC) with digital de-waxing; another outlier removal. The deep learning model was a convolutional neural network (CNN) fused with a fully connected neural network (FCNN). The CNN was built with 2 Conv1D-ReLU-MaxPooling1D-Dropout layers. The kernel size was set to 5 and dropout of 0.5. Dense layers were built by two layers of neurons-BatchNorm-ReLU-Dropout, with 100 and 50 neurons, dropout of 0.2. The output was a single neuron with sigmoid activation. Binary cross-entropy loss function was adopted with Adam optimizer. Accuracy metric was calculated during the training, where a threshold of 0.5 was applied on the output predictions. Model was trained by a 4-fold cross-validation by 20 epochs and using a batch size of 250. The train accuracy was 0.978/0.004 (mean/std), while the testing accuracy was 0.969/0.008, demonstrating a generalized model without overfitting. Accuracies near one indicate the proposed model as a potential technique for the breast cancer vs adenosis classification, where hyperparameters and the architecture should be optimized along higher sample number acquisition.
Como referenciar
DEL-VALLE, MATHEUS; SANTOS, MOISES O. dos; SANTOS, SOFIA N. dos; BERNARDES, EMERSON S.; ZEZELL, DENISE M. A deep learning approach for breast tissue malignancy diagnosis using micro-FTIR hyperspectral imaging. In: ENCONTRO DE OUTONO DA SOCIEDADE BRASILEIRA DE FÍSICA, 44., 21-25 de junho, 2021, Online. Resumo... São Paulo, SP: Sociedade Brasileira de Física, 2021. Disponível em: http://repositorio.ipen.br/handle/123456789/32690. Acesso em: 26 Apr 2024.
Esta referência é gerada automaticamente de acordo com as normas do estilo IPEN/SP (ABNT NBR 6023) e recomenda-se uma verificação final e ajustes caso necessário.