Abstract

Traditional clinician diagnosis, which requires intensive manual effort from experienced medical doctors and radiologists, is notoriously time-consuming, costly and at times error prone. To alleviate these issues, computer-aided diagnosis systems are often used to improve accuracy in early detection, diagnosis, treatment plan and an outcome prediction. While these systems are making strides, significant challenges remain due the scarcity of publicly available data, high annotation cost, and suboptimal performance in detecting rare targets. In this thesis, we design efficient deep learning models for computer-aided medical diagnosis. The contributions are two-fold: First, we introduce an over-sampling method for learning the inter-class mapping between under-represented class samples and over-represented samples in a bid to generate under-represented class samples using unpaired image-to-image translation. These synthetic images are then used as additional training data in the task of detecting abnormalities (i.e. melanoma, COVID-19). Our method achieves improved performance on a standard skin lesion classification benchmark. We show through feature visualization that our approach leads to context based lesion assessment that can reach an expert dermatologist level. Additional experiments also demonstrate the effectiveness of our model in COVID-19 detection from chest radiography images. The synthetic images not only improve performance of various deep learning architectures when used as additional training data under heavy imbalance conditions, but also detect the target class with high confidence.

Second, we present a simple, yet effective end-to-end depthwise encoder-decoder fully convolutional network architecture, dubbed Sharp U-Net, for binary and multi-class biomedical image segmentation. Instead of applying a plain skip connection such as U-Net, a depthwise convolution of the encoder feature map with a sharpening kernel filter is employed prior to merging the encoder and decoder features, thereby producing a sharpened intermediate feature map of the same size as the encoder map. Using this sharpening filter layer, we are able to not only fuse semantically less dissimilar features, but also smooth out artifacts throughout the network layers during the early stages of training. Our extensive experiments on six datasets show that the proposed Sharp U-Net model consistently outperforms or matches the recent state-of-the-art baselines in both binary and multi-class segmentation tasks, while adding no extra learnable parameters.