Toward Learning Trustworthily from Data Combining Privacy, Fairness, and Explainability: An Application to Face Recognition.

In many decision-making scenarios, ranging from recreational activities to healthcare and policing, the use of artificial intelligence coupled with the ability to learn from historical data is becoming ubiquitous. This widespread adoption of automated systems is accompanied by the increasing concerns regarding their ethical implications. Fundamental rights, such as the ones that require the preservation of privacy, do not discriminate based on sensible attributes (e.g., gender, ethnicity, political/sexual orientation), or require one to provide an explanation for a decision, are daily undermined by the use of increasingly complex and less understandable yet more accurate learning algorithms. For this purpose, in this work, we work toward the development of systems able to ensure trustworthiness by delivering privacy, fairness, and explainability by design. In particular, we show that it is possible to simultaneously learn from data while preserving the privacy of the individuals thanks to the use of Homomorphic Encryption, ensuring fairness by learning a fair representation from the data, and ensuring explainable decisions with local and global explanations without compromising the accuracy of the final models. We test our approach on a widespread but still controversial application, namely face recognition, using the recent FairFace dataset to prove the validity of our approach.

Deciphering multiple sclerosis disability with deep learning attention maps on clinical MRI.

The application of convolutional neural networks (CNNs) to MRI data has emerged as a promising approach to achieving unprecedented levels of accuracy when predicting the course of neurological conditions, including multiple sclerosis, by means of extracting image features not detectable through conventional methods. Additionally, the study of CNN-derived attention maps, which indicate the most relevant anatomical features for CNN-based decisions, has the potential to uncover key disease mechanisms leading to disability accumulation. From a cohort of patients prospectively followed up after a first demyelinating attack, we selected those with T1-weighted and T2-FLAIR brain MRI sequences available for image analysis and a clinical assessment performed within the following six months (N = 319). Patients were divided into two groups according to expanded disability status scale (EDSS) score: ≥3.0 and < 3.0. A 3D-CNN model predicted the class using whole-brain MRI scans as input. A comparison with a logistic regression (LR) model using volumetric measurements as explanatory variables and a validation of the CNN model on an independent dataset with similar characteristics (N = 440) were also performed. The layer-wise relevance propagation method was used to obtain individual attention maps. The CNN model achieved a mean accuracy of 79% and proved to be superior to the equivalent LR-model (77%). Additionally, the model was successfully validated in the independent external cohort without any re-training (accuracy = 71%). Attention-map analyses revealed the predominant role of frontotemporal cortex and cerebellum for CNN decisions, suggesting that the mechanisms leading to disability accrual exceed the mere presence of brain lesions or atrophy and probably involve how damage is distributed in the central nervous system.

Interpretable deep learning for the prognosis of long-term functional outcome post-stroke using acute diffusion weighted imaging.

Advances in deep learning can be applied to acute stroke imaging to build powerful and explainable prediction models that could supersede traditionally used biomarkers. We aimed to evaluate the performance and interpretability of a deep learning model based on convolutional neural networks (CNN) in predicting long-term functional outcome with diffusion-weighted imaging (DWI) acquired at day 1 post-stroke. Ischemic stroke patients (n = 322) were included from the ASTER and INSULINFARCT trials as well as the Pitié-Salpêtrière registry. We trained a CNN to predict long-term functional outcome assessed at 3 months with the modified Rankin Scale (dichotomized as good [mRS ≤ 2] vs. poor [mRS ≥ 3]) and compared its performance to two logistic regression models using lesion volume and ASPECTS. The CNN contained an attention mechanism, which allowed to visualize the areas of the brain that drove prediction. The deep learning model yielded a significantly higher area under the curve (0.83 95%CI [0.78-0.87]) than lesion volume (0.78 [0.73-0.83]) and ASPECTS (0.77 [0.71-0.83]) (p < 0.05). Setting all classifiers to the specificity as the deep learning model (i.e., 0.87 [0.82-0.92]), the CNN yielded a significantly higher sensitivity (0.67 [0.59-0.73]) than lesion volume (0.48 [0.40-0.56]) and ASPECTS (0.50 [0.41-0.58]) (p = 0.002). The attention mechanism revealed that the network learned to naturally attend to the lesion to predict outcome.

Attention-Based Deep Learning System for Classification of Breast Lesions-Multimodal, Weakly Supervised Approach.

Breast cancer is the most frequent female cancer, with a considerable disease burden and high mortality. Early diagnosis with screening mammography might be facilitated by automated systems supported by deep learning artificial intelligence. We propose a model based on a weakly supervised Clustering-constrained Attention Multiple Instance Learning (CLAM) classifier able to train under data scarcity effectively. We used a private dataset with 1174 non-cancer and 794 cancer images labelled at the image level with pathological ground truth confirmation. We used feature extractors (ResNet-18, ResNet-34, ResNet-50 and EfficientNet-B0) pre-trained on ImageNet. The best results were achieved with multimodal-view classification using both CC and MLO images simultaneously, resized by half, with a patch size of 224 px and an overlap of 0.25. It resulted in AUC-ROC = 0.896 ± 0.017, F1-score 81.8 ± 3.2, accuracy 81.6 ± 3.2, precision 82.4 ± 3.3, and recall 81.6 ± 3.2. Evaluation with the Chinese Mammography Database, with 5-fold cross-validation, patient-wise breakdowns, and transfer learning, resulted in AUC-ROC 0.848 ± 0.015, F1-score 78.6 ± 2.0, accuracy 78.4 ± 1.9, precision 78.8 ± 2.0, and recall 78.4 ± 1.9. The CLAM algorithm's attentional maps indicate the features most relevant to the algorithm in the images. Our approach was more effective than in many other studies, allowing for some explainability and identifying erroneous predictions based on the wrong premises.

Explaining a Deep Learning Based Breast Ultrasound Image Classifier with Saliency Maps.

Aim of the study: Deep neural networks have achieved good performance in breast mass classification in ultrasound imaging. However, their usage in clinical practice is still limited due to the lack of explainability of decisions conducted by the networks. In this study, to address the explainability problem, we generated saliency maps indicating ultrasound image regions important for the network's classification decisions. Material and methods: Ultrasound images were collected from 272 breast masses, including 123 malignant and 149 benign. Transfer learning was applied to develop a deep network for breast mass classification. Next, the class activation mapping technique was used to generate saliency maps for each image. Breast mass images were divided into three regions: the breast mass region, the peritumoral region surrounding the breast mass, and the region below the breast mass. The pointing game metric was used to quantitatively assess the overlap between the saliency maps and the three selected US image regions. Results: Deep learning classifier achieved the area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity of 0.887, 0.835, 0.801, and 0.868, respectively. In the case of the correctly classified test US images, analysis of the saliency maps revealed that the decisions of the network could be associated with the three selected regions in 71% of cases. Conclusions: Our study is an important step toward better understanding of deep learning models developed for breast mass diagnosis. We demonstrated that the decisions made by the network can be related to the appearance of certain tissue regions in breast mass US images.

Multi-scale information with attention integration for classification of liver fibrosis in B-mode US image.

BACKGROUND AND OBJECTIVE: Chronic hepatitis B (CHB) is one of the most common liver diseases in the world, which threats a lot to people's usual life. The increased deposition of fibrotic tissues in livers for patients with CHB may lead to the development of liver cirrhosis, hepatocellular carcinoma, or even liver failure. Accurate fibrosis staging is very important for the targeted treatment of liver fibrosis and its recovery. METHODS: In this paper, we propose a new deep convolutional neural network (DCNN) with functions of multi-scale information extraction and attention integration for more accurate liver fibrosis classification from ultrasound (US) images. The proposed network uses two pyramid-structured CNN elements to extract multi-scale features from US images. Such a design significantly enlarges the receptive field of the convolution layer, such that more useful information can be explored by the neural network to associate with the final classification. Based on this, a new feature distillation method is also proposed to enhance the ability of deep features derived from multi-scale information. The proposed distillation method employs attention maps to automatically extract class-related features from multi-scale information, which effectively suppress the influence of potential distractors. RESULTS: Experimental results on the US liver fibrosis dataset collected from 286 participants show that the proposed deep framework achieves promising classification performance. The proposed method achieves a classification accuracy of 95.66% on the test dataset. CONCLUSION: Our proposed framework could stage liver fibrosis highly accurately. It might provide effective suggestions for the clinical treatment of liver fibrosis that can facilitate its recovery.

OTNet: A CNN Method Based on Hierarchical Attention Maps for Grading Arteriosclerosis of Fundus Images with Small Samples.

The severity of fundus arteriosclerosis can be determined and divided into four grades according to fundus images. Automatically grading of the fundus arteriosclerosis is helpful in clinical practices, so this paper proposes a convolutional neural network (CNN) method based on hierarchical attention maps to solve the automatic grading problem. First, we use the retinal vessel segmentation model to separate the important vascular region and the non-vascular background region from the fundus image and obtain two attention maps. The two maps are regarded as inputs to construct a two-stream CNN (TSNet), to focus on feature information through mutual reference between the two regions. In addition, we use convex hull attention maps in the one-stream CNN (OSNet) to learn valuable areas where the retinal vessels are concentrated. Then, we design an integrated OTNet model which is composed of TSNet that learns image feature information and OSNet that learns discriminative areas. After obtaining the representation learning parts of the two networks, we can train the classification layer to achieve better results. Our proposed TSNet reaches the AUC value of 0.796 and the ACC value of 0.592 on the testing set, and the integrated model OTNet reaches the AUC value of 0.806 and the ACC value of 0.606, which are better than the results of other comparable models. As far as we know, this is the first attempt to use deep learning to classify the severity of atherosclerosis in fundus images. The prediction results of our proposed method can be accepted by doctors, which shows that our method has a certain application value.