RESUMEN
BACKGROUND AND OBJECTIVE: Training a deep convolutional neural network (CNN) for automatic image classification requires a large database with images of labeled samples. However, in some applications such as biology and medicine only a few experts can correctly categorize each sample. Experts are able to identify small changes in shape and texture which go unnoticed by untrained people, as well as distinguish between objects in the same class that present drastically different shapes and textures. This means that currently available databases are too small and not suitable to train deep learning models from scratch. To deal with this problem, data augmentation techniques are commonly used to increase the dataset size. However, typical data augmentation methods introduce artifacts or apply distortions to the original image, which instead of creating new realistic samples, obtain basic spatial variations of the original ones. METHODS: We propose a novel data augmentation procedure which generates new realistic samples, by combining two samples that belong to the same class. Although the idea behind the method described in this paper is to mimic the variations that diatoms experience in different stages of their life cycle, it has also been demonstrated in glomeruli and pollen identification problems. This new data augmentation procedure is based on morphing and image registration methods that perform diffeomorphic transformations. RESULTS: The proposed technique achieves an increase in accuracy over existing techniques of 0.47%, 1.47%, and 0.23% for diatom, glomeruli and pollen problems respectively. CONCLUSIONS: For the Diatom dataset, the method is able to simulate the shape changes in different diatom life cycle stages, and thus, images generated resemble newly acquired samples with intermediate shapes. In fact, the other methods compared obtained worse results than those which were not using data augmentation. For the Glomeruli dataset, the method is able to add new samples with different shapes and degrees of sclerosis (through different textures). This is the case where our proposed DA method is more beneficial, when objects highly differ in both shape and texture. Finally, for the Pollen dataset, since there are only small variations between samples in a few classes and this dataset has other features such as noise which are likely to benefit other existing DA techniques, the method still shows an improvement of the results.
Asunto(s)
Manejo de Datos , Redes Neurales de la Computación , Bases de Datos Factuales , HumanosRESUMEN
Reliable counting of glomeruli and evaluation of glomerulosclerosis in renal specimens are essential steps to assess morphological changes in kidney and identify individuals requiring treatment. Because microscopic identification of sclerosed glomeruli performed under the microscope is labor intensive, we developed a deep learning (DL) approach to identify and classify glomeruli as normal or sclerosed in digital whole slide images (WSIs). The segmentation and classification of glomeruli was performed by the U-Net model. Subsequently, glomerular classifications were refined based on glomerular histomorphometry. The U-Net model was trained using patches from Periodic Acid-Schiff (PAS) stained WSIs (n=31) from the AIDPATH - a multi-center dataset, and then tested on an independent set of WSIs (n=20) including PAS (n=6), and hematoxylin and eosin (H&E) stained WSIs (n=14) from four other institutions. The training and test WSIs were obtained from formalin fixed and paraffin embedded blocks with of human kidney specimens each presenting various proportions of normal and sclerosed glomeruli. In the PAS stained WSIs, normal and sclerosed glomeruli were respectively classified with the F1-score of 97.5% and 68.8%. In the H&E stained WSIs, the F1-scores of 90.8% and 78.1% were achieved. Regardless the tissue staining, the glomeruli in the test WSIs were classified with the F1-score of 94.5% (n=923, normal) and 76.8% for (n=261, sclerosed). These results demonstrate for the first time that a framework based on the U-Net model trained with glomerular patches from PAS stained WSIs can reliably segment and classify normal and sclerosed glomeruli in PAS and also H&E stained WSIs. Our approach yielded higher accuracy of glomerular classifications than some of the recently published methods. Additionally, our test set of images with ground truth is publicly available.