Abstract

Abstract
Diabetic macular edema is a leading cause of visual loss for patients with diabetes. While diagnosis can only be performed by optical coherence tomography, diabetic macular edema risk assessment is often performed in eye fundus images in screening scenarios through the detection of hard exudates. Such screening scenarios are often associated with large amounts of data, high costs and high burden on specialists, motivating then the development of methodologies for automatic diabetic macular edema risk prediction. Nevertheless, significant dataset domain bias, due to different acquisition equipment, protocols and/or different populations can have significantly detrimental impact on the performance of automatic methods when transitioning to a new dataset, center or scenario. As such, in this study, a method based on residual neural networks is proposed for the classification of diabetic macular edema risk. This method is then validated across multiple public datasets, simulating the deployment in a multi-center setting and thereby studying the method’s generalization capability and existing dataset domain bias. Furthermore, the method is tested on a private dataset which more closely represents a realistic screening scenario. An average area under the curve across all public datasets of 0.891 ± 0.013 was obtained with a ResNet50 architecture trained on a limited amount of images from a single public dataset (IDRiD). It is also shown that screening scenarios are significantly more challenging and that training across multiple datasets leads to an improvement of performance (area under the curve of 0.911 ± 0.009).

Abstract
Diabetic retinopathy (DR) is a diabetes complication, which in extreme situations may lead to blindness. Since the first stages are often asymptomatic, regular eye examinations are required for an early diagnosis. As microaneurysms (MAs) are one of the first signs of DR, several automated methods have been proposed for their detection in order to reduce the ophthalmologists' workload. Although local convergence filters (LCFs) have already been applied for feature extraction, their potential as MA enhancement operators was not explored yet. In this work, we propose a sliding band filter for MA enhancement aiming at obtaining a set of initial MA candidates. Then, a combination of the filter responses with color, contrast and shape information is used by an ensemble of classifiers for final candidate classification. Finally, for each eye fundus image, a score is computed from the confidence values assigned to the MAs detected in the image. The performance of the proposed methodology was evaluated in four datasets. At the lesion level, sensitivities of 64% and 81% were achieved for an average of 8 false positives per image (FPIs) in e-ophtha MA and SCREEN-DR, respectively. In the last dataset, an AUC of 0.83 was also obtained for DR detection.

Abstract
Proliferative diabetic retinopathy (PDR) is an advanced diabetic retinopathy stage, characterized by neovascularization, which leads to ocular complications and severe vision loss. However, the available DR-labeled retinal image datasets have a small representation of images of the severest DR grades, and thus there is lack of PDR cases for training DR grading models. Additionally, the criteria for labelling these images in the publicly available datasets is not always clear, with some images which do not show typical PDR lesions being labeled as PDR due to the presence of photo-coagulation treatment and laser marks. This problem, together with the datasets' high class imbalance, leads to a limited variability of the samples, which the typical data augmentation and class balancing cannot fully mitigate. We propose a heuristic-based data augmentation scheme based on the synthesis of neovessel (NV)-like structures that compensates for the lack of PDR cases in DR-labeled datasets. The proposed neovessel generation algorithm relies on the general knowledge of common location and shape of these structures. NVs are generated and introduced in pre-existent retinal images which can then be used for enlarging deep neural networks' training sets. The data augmentation scheme was tested on multiple datasets, and allows to improve the model's capacity to detect NVs.

Abstract
Lung cancer is the deadliest type of cancer worldwide and late detection is the major factor for the low survival rate of patients. Low dose computed tomography has been suggested as a potential screening tool but manual screening is costly and time-consuming. This has fuelled the development of automatic methods for the detection, segmentation and characterisation of pulmonary nodules. In spite of promising results, the application of automatic methods to clinical routine is not straightforward and only a limited number of studies have addressed the problem in a holistic way. With the goal of advancing the state of the art, the Lung Nodule Database (LNDb) Challenge on automatic lung cancer patient management was organized. The LNDb Challenge addressed lung nodule detection, segmentation and characterization as well as prediction of patient follow-up according to the 2017 Fleischner society pulmonary nodule guidelines. 294 CT scans were thus collected retrospectively at the Centro Hospitalar e Universitrio de So Joo in Porto, Portugal and each CT was annotated by at least one radiologist. Annotations comprised nodule centroids, segmentations and subjective characterization. 58 CTs and the corresponding annotations were withheld as a separate test set. A total of 947 users registered for the challenge and 11 successful submissions for at least one of the sub-challenges were received. For patient follow-up prediction, a maximum quadratic weighted Cohen's kappa of 0.580 was obtained. In terms of nodule detection, a sensitivity below 0.4 (and 0.7) at 1 false positive per scan was obtained for nodules identified by at least one (and two) radiologist(s). For nodule segmentation, a maximum Jaccard score of 0.567 was obtained, surpassing the interobserver variability. In terms of nodule texture characterization, a maximum quadratic weighted Cohen's kappa of 0.733 was obtained, with part solid nodules being particularly challenging to classify correctly. Detailed analysis of the proposed methods and the differences in performance allow to identify the major challenges remaining and future directions data collection, augmentation/generation and evaluation of under-represented classes, the incorporation of scan-level information for better decision making and the development of tools and challenges with clinical-oriented goals. The LNDb Challenge and associated data remain publicly available so that future methods can be tested and benchmarked, promoting the development of new algorithms in lung cancer medical image analysis and patient followup recommendation.

Abstract
Current state-of-the-art medical image segmentation methods require high quality datasets to obtain good performance. However, medical specialists often disagree on diagnosis, hence, datasets contain contradictory annotations. This, in turn, leads to difficulties in the optimization process of Deep Learning models and hinder performance. We propose a method to estimate uncertainty in Convolutional Neural Network (CNN) segmentation models, that makes the training of CNNs more robust to contradictory annotations. In this work, we model two types of uncertainty, heteroscedastic and epistemic, without adding any additional supervisory signal other than the ground-truth segmentation mask. As expected, the uncertainty is higher closer to vessel boundaries, and on top of thinner and less visible vessels where it is more likely for medical specialists to disagree. Therefore, our method is more suitable to learn from datasets created with heterogeneous annotators. We show that there is a correlation between the uncertainty estimated by our method and the disagreement in the segmentation provided by two different medical specialists. Furthermore, by explicitly modeling the uncertainty, the Intersection over Union of the segmentation network improves 5.7 percentage points.