| dc.creator | VALLE, MATHEUS del | |
| dc.creator | SANTOS, MOISES O. dos | |
| dc.creator | CONGRESS OF THE INTERNATIONAL UNION FOR PURE APPLIED BIOPHYSICS, 20th; ANNUAL MEETING OF THE BRAZILIAN SOCIETY FOR BIOCHEMISTRY AND MOLECULAR BIOLOGY, 50th; CONGRESS OF BRAZILIAN BIOPHYSICS SOCIETY, 45th; BRAZILIAN SOCIETY ON NUCLEAR BIOSCIENCES CONGRESS, 13th | |
| dc.date | 2022-03-25T15:36:05Z | |
| dc.date | 2022-03-25T15:36:05Z | |
| dc.date | October 4-8, 2021 | |
| dc.date.accessioned | 2023-09-28T14:21:36Z | |
| dc.date.available | 2023-09-28T14:21:36Z | |
| dc.identifier | http://repositorio.ipen.br/handle/123456789/32862 | |
| dc.identifier.uri | https://repositorioslatinoamericanos.uchile.cl/handle/2250/9003081 | |
| dc.description | The estimative of new breast cancer cases was of 2.1 million of new breast cancer cases in 2018, hence
being the most incident type of cancer in women. The improvement of its diagnosis has been the aim of
many researchers, including vibrational spectroscopy teams. With the advancement of the artificial
intelligence, a field of computer science to enhance intelligence into computer systems, specially of the
deep learning, big data acquired from spectroscopy image has entered a new era. Therefore, the proposal
of this work was to diagnose breast tissue samples as malignant (cancer) or benign (adenosis) using
deep learning techniques. Micro-FTIR spectroscopy images were acquired from BR804b human breast
tissue microarray (Biomax, USA), resulting in more than 100 thousand spectra for each group. A k-means
approach was established to separate spectra into three clusters: tissue, paraffin and slide. A
preprocessing step was applied by the following pipeline: outlier removal; biofingerprint truncation;
Savitzky???Golay filter to smooth and to obtain the second derivative; extended multiplicative signal
correction to correct spectra and remove the paraffin contribution. The deep learning algorithm was built
using two-layers of one-dimensional convolutional neural network (CNN) connected to a two-layers (100
and 50 neurons) feedforward network (FFN). Both networks used dropout layers of 50% and rectified
linear unit activations. CNN kernel size was set to 5. The output neuron used a sigmoid activation. Adam
optimizer was applied to train the networks, using a binary cross-entropy loss to improve the weights. A
4-fold cross-validation of 20 epochs and batch size of 250 was performed. The networks exhibited an
accuracy of (97.8 ?? 0.4)% during the training stage, and (96.9 ?? 0.8)% during the testing stage,
demonstrating a generalized classification. Accuracies of almost 100% indicates this approach as a
potential technique for the breast diagnosis. | |
| dc.format | 42-42 | |
| dc.publisher | Sociedade Brasileira de Bioqu??mica e Biologia Molecular (SBBq) | |
| dc.rights | openAccess | |
| dc.subject | fourier transformation | |
| dc.subject | infrared spectrometers | |
| dc.subject | artificial intelligence | |
| dc.subject | mammary glands | |
| dc.subject | neoplasms | |
| dc.title | Breast tissue diagnosis using artificial intelligence applied to FTIR spectroscopy images | |
| dc.type | Resumo de eventos cient??ficos | |
| dc.coverage | I | |
| dc.local | S??o Paulo, SP | |
