In this paper, for the first time, we propose an evaluation method for deep learning models that assesses the performance of a model not only in an unseen test scenario, but also in extreme cases of noise, outliers and ambiguous input data. To this end, we utilize adversarial examples, images that fool machine learning models, while looking imperceptibly different from original data, as a measure to evaluate the robustness of a variety of medical imaging models. Through extensive experiments on skin lesion classification and whole brain segmentation with state-of-the-art networks such as Inception and UNet, we show that models that achieve comparable performance regarding generalizability may have significant variations in their perception of the underlying data manifold, leading to an extensive performance gap in their robustness.
Within medical imaging, manual curation of sufficient well-labeled samples is cost, time and scale-prohibitive. To improve the representativeness of the training dataset, for the first time, we present an approach to utilize large amounts of freely available web data through web-crawling. To handle noise and weak nature of web annotations, we propose a two-step transfer learning based training process with a robust loss function, termed as Webly Supervised Learning (WSL) to train deep models for the task. We also leverage search by image to improve the search specificity of our web-crawling and reduce cross-domain noise. Within WSL, we explicitly model the noise structure between classes and incorporate it to selectively distill knowledge from the web data during model training. To demonstrate improved performance due to WSL, we benchmarked on a publicly available 10-class fine-grained skin lesion classification dataset and report a significant improvement of top-1 classification accuracy from 71.25 % to 80.53 % due to the incorporation of web-supervision.
Deep learning algorithms, especially convolutional neural networks, have become a methodology of choice in medical image analysis. However, recent studies in computer vision show that even a small modification of input image intensities may cause a deep learning model to classify the image differently. In medical imaging, the distribution of image intensities is related to applied image reconstruction algorithm. In this paper we investigate the impact of ultrasound image reconstruction method on breast lesion classification with neural transfer learning. Due to high dynamic range raw ultrasonic signals are commonly compressed in order to reconstruct B-mode images. Based on raw data acquired from breast lesions, we reconstruct B-mode images using different compression levels. Next, transfer learning is applied for classification. Differently reconstructed images are employed for training and evaluation. We show that the modification of the reconstruction algorithm leads to decrease of classification performance. As a remedy, we propose a method of data augmentation. We show that the augmentation of the training set with differently reconstructed B-mode images leads to a more robust and efficient classification. Our study suggests that it is important to take into account image reconstruction algorithms implemented in medical scanners during development of computer aided diagnosis systems.
Generative Adversarial Networks (GANs) have been successfully used to synthesize realistically looking images of faces, scenery and even medical images. Unfortunately, they usually require large training datasets, which are often scarce in the medical field, and to the best of our knowledge GANs have been only applied for medical image synthesis at fairly low resolution. However, many state-of-the-art machine learning models operate on high resolution data as such data carries indispensable, valuable information. In this work, we try to generate realistically looking high resolution images of skin lesions with GANs, using only a small training dataset of 2000 samples. The nature of the data allows us to do a direct comparison between the image statistics of the generated samples and the real dataset. We both quantitatively and qualitatively compare state-of-the-art GAN architectures such as DCGAN and LAPGAN against a modification of the latter for the task of image generation at a resolution of 256x256px. Our investigation shows that we can approximate the real data distribution with all of the models, but we notice major differences when visually rating sample realism, diversity and artifacts. In a set of use-case experiments on skin lesion classification, we further show that we can successfully tackle the problem of heavy class imbalance with the help of synthesized high resolution melanoma samples.
In many machine learning applications, from medical diagnostics to autonomous driving, the availability of prior knowledge can be used to improve the predictive performance of learning algorithms and incorporate `physical,' `domain knowledge,' or `common sense' concepts into training of machine learning systems as well as verify constraints/properties of the systems. We explore the learning using privileged information paradigm and show how to incorporate the privileged information, such as segmentation mask available along with the classification label of each example, into the training stage of convolutional neural networks. This is done by augmenting the CNN model with an architectural component that effectively focuses model's attention on the desired region of the input image during the training process and that is transparent to the network's label prediction mechanism at testing. This component effectively corresponds to the visualization strategy for identifying the parts of the input, often referred to as visualization mask, that most contribute to the prediction, yet uses this strategy in reverse to the classical setting in order to enforce the desired visualization mask instead. We verify our proposed algorithms through exhaustive experiments on benchmark ImageNet and PASCAL VOC data sets and achieve improvements in the performance of $2.4\%$ and $2.7\%$ over standard single-supervision model training. Finally, we confirm the effectiveness of our approach on skin lesion classification problem.