Loading...
3 results
Search Results
Now showing 1 - 3 of 3
- Lesion Classification in Mammograms Using Convolutional Neural Networks and Transfer LearningPublication . Perre, Ana Catarina; Alexandre, Luís; Freire, Luís C.Computer-Aided Detection/Diagnosis (CAD) tools were created to assist the detection and diagnosis of early stage cancers, decreasing false negative rate and improving radiologists’ efficiency. Convolutional Neural Networks (CNNs) are one example of deep learning algorithms that proved to be successful in image classification. In this paper we aim to study the application of CNNs to the classification of lesions in mammograms. One major problem in the training of CNNs for medical applications is the large dataset of images that is often required but seldom available. To solve this problem, we use a transfer learning approach, wich is based on three different networks that were pre-trained on the Imagenet dataset. We then investigate the performance of these pre-trained CNNs and two types of image normalization to classify lesions in mammograms. The best results were obtained using the Caffe reference model for the CNN with no image normalization.
- Image Normalization Influence in Mammographic Classification with CNNsPublication . Perre, Ana Catarina; Alexandre, Luís; Freire, Luís C.In order to improve the performance of Convolutional Neural Networks (CNN) in the classification of mammographic images, many researchers choose to apply a normalization method during the pre-processing stage. In this work, we aimed to assess the impact of 6 different normalization methods in the classification performance of 2 CNNs. We have also explored 5 classifiers, being the first one the CNN itself. The other 4 correspond to Support Vector Machine (SVM), Random Forest (RF), Simple Logistic (SL) and Voted Perceptron (VP) classifiers, all of them fed with features extracted from one of the layers - comprised between the sixteenth and the nineteenth - of the CNN. The last 3 classifiers were tested with different options for data testing presentation, according to theWeka software: Supplied Test Set (STS), 10-fold Cross Validation (10-FCV) and Percentage Split (PS). Results indicate that the effect of image normalization in the performance of the CNNs depends on which network is chosen to make the classification; besides, the normalization method that seems to have the most positive impact is the one that subtracts to each image the corresponding image mean and divide it by the standard deviation (best AUC mean values were 0.786 for CNN-F and 0.790 for Caffe; the best run AUC values were, respectively, 0.793 and 0.791. Layer 1 freezing decreased the running time and did not harm the classification performance. Regarding the different classifiers, CNNs used alone with softmax yielded the best results, with the exception of the RF and SL classifiers, both using the 10-FCV and PS options; however, with these options, we cannot guarantee that the test set images are presented for the first time to the network.
- Lesion classification in mammograms using convolutional neural networks and transfer learningPublication . Perre, Ana Catarina; Alexandre, Luís; Freire, Luís C.Convolutional neural networks (CNNs) have recently been successfully used in the medical field to detect and classify pathologies in different imaging modalities, including in mammography. One disadvantage of CNNs is the need for large training datasets, which are particularly difficult to obtain in the medical domain. One way to solve this problem is using a transfer learning approach, in which a CNN, previously pre-trained with a large amount of labelled non-medical data, is subsequently finetuned using a smaller dataset of medical data. In this paper, we use such a transfer learning approach, which is applied to three different networks that were pre-trained using the Imagenet dataset. We investigate how the performance of these pre-trained CNNs to classify lesions in mammograms is affected by the use, or not, of normalised images during the fine-tuning stage. We also assess the performance of a support vector machine fed with features extracted from the CNN and the combined use of handcrafted features to complement the CNN-extracted features. The obtained results are encouraging.