Generative models are popular for medical imaging tasks such as anomaly detection, feature extraction, data visualization, or image generation. Since they are parameterized by deep learning models, they are often sensitive to distribution shifts and unreliable when applied to out-of-distribution data, creating a risk of, e.g. underrepresentation bias. This behavior can be flagged using uncertainty quantification methods for generative models, but their availability remains limited. We propose SLUG: A new UQ method for VAEs that combines recent advances in Laplace approximations with stochastic trace estimators to scale gracefully with image dimensionality. We show that our UQ score -- unlike the VAE's encoder variances -- correlates strongly with reconstruction error and racial underrepresentation bias for dermatological images. We also show how pixel-wise uncertainty can detect out-of-distribution image content such as ink, rulers, and patches, which is known to induce learning shortcuts in predictive models.

Ewing's sarcoma (ES), characterized by a high density of small round blue cells without structural organization, presents a significant health concern, particularly among adolescents aged 10 to 19. Artificial intelligence-based systems for automated analysis of histopathological images are promising to contribute to an accurate diagnosis of ES. In this context, this study explores the feature extraction ability of different pre-training strategies for distinguishing ES from other soft tissue or bone sarcomas with similar morphology in digitized tissue microarrays for the first time, as far as we know. Vision-language supervision (VLS) is compared to fully-supervised ImageNet pre-training within a multiple instance learning paradigm. Our findings indicate a substantial improvement in diagnostic accuracy with the adaption of VLS using an in-domain dataset. Notably, these models not only enhance the accuracy of predicted classes but also drastically reduce the number of trainable parameters and computational costs.

Domain shift in the field of histopathological imaging is a common phenomenon due to the intra- and inter-hospital variability of staining and digitization protocols. The implementation of robust models, capable of creating generalized domains, represents a need to be solved. In this work, a new domain adaptation method to deal with the variability between histopathological images from multiple centers is presented. In particular, our method adds a training constraint to the supervised contrastive learning approach to achieve domain adaptation and improve inter-class separability. Experiments performed on domain adaptation and classification of whole-slide images of six skin cancer subtypes from two centers demonstrate the method's usefulness. The results reflect superior performance compared to not using domain adaptation after feature extraction or staining normalization.

Vision-language supervision has made remarkable strides in learning visual representations from textual guidance. In digital pathology, vision-language models (VLM), pre-trained on curated datasets of histological image-captions, have been adapted to downstream tasks, such as region of interest classification. Zero-shot transfer for slide-level prediction has been formulated by MI-Zero [1], but it exhibits high variability depending on the textual prompts. Inspired by prototypical learning, we propose MI-VisionShot, a training-free adaptation method on top of VLMs to predict slide-level labels in few-shot learning scenarios. Our framework takes advantage of the excellent representation learning of VLM to create prototype-based classifiers under a multipleinstance setting by retrieving the most discriminative patches within each slide. Experimentation through different settings shows the ability of MI-VisionShot to surpass zero-shot transfer with lower variability, even in low-shot scenarios.

Natural databases for Speech Emotion Recognition (SER) are scarce and often rely on staged scenarios, such as films or television shows, limiting their application in real-world contexts. We developed and publicly released the Emotional Voice Messages (EMOVOME) database, including 999 voice messages from real conversations of 100 Spanish speakers on a messaging app, labeled in continuous and discrete emotions by expert and non-expert annotators. We evaluated speaker-independent SER models using a standard set of acoustic features and transformer-based models. We compared the results with reference databases including acted and elicited speech, and analyzed the influence of annotators and gender fairness. The pre-trained UniSpeech-SAT-Large model achieved the highest results, 61.64% and 55.57% Unweighted Accuracy (UA) for 3-class valence and arousal prediction respectively on EMOVOME, a 10% improvement over baseline models. For the emotion categories, 42.58% UA was obtained. EMOVOME performed lower than the acted RAVDESS database. The elicited IEMOCAP database also outperformed EMOVOME in predicting emotion categories, while similar results were obtained in valence and arousal. EMOVOME outcomes varied with annotator labels, showing better results and fairness when combining expert and non-expert annotations. This study highlights the gap between staged and real-life scenarios, supporting further advancements in recognizing genuine emotions.

Emotional Voice Messages (EMOVOME) is a spontaneous speech dataset containing 999 audio messages from real conversations on a messaging app from 100 Spanish speakers, gender balanced. Voice messages were produced in-the-wild conditions before participants were recruited, avoiding any conscious bias due to laboratory environment. Audios were labeled in valence and arousal dimensions by three non-experts and two experts, which were then combined to obtain a final label per dimension. The experts also provided an extra label corresponding to seven emotion categories. To set a baseline for future investigations using EMOVOME, we implemented emotion recognition models using both speech and audio transcriptions. For speech, we used the standard eGeMAPS feature set and support vector machines, obtaining 49.27% and 44.71% unweighted accuracy for valence and arousal respectively. For text, we fine-tuned a multilingual BERT model and achieved 61.15% and 47.43% unweighted accuracy for valence and arousal respectively. This database will significantly contribute to research on emotion recognition in the wild, while also providing a unique natural and freely accessible resource for Spanish.

We present a pioneering investigation into the application of deep learning techniques to analyze histopathological images for addressing the substantial challenge of automated prognostic prediction. Prognostic prediction poses a unique challenge as the ground truth labels are inherently weak, and the model must anticipate future events that are not directly observable in the image. To address this challenge, we propose a novel three-part framework comprising of a convolutional network based tissue segmentation algorithm for region of interest delineation, a contrastive learning module for feature extraction, and a nested multiple instance learning classification module. Our study explores the significance of various regions of interest within the histopathological slides and exploits diverse learning scenarios. The pipeline is initially validated on artificially generated data and a simpler diagnostic task. Transitioning to prognostic prediction, tasks become more challenging. Employing bladder cancer as use case, our best models yield an AUC of 0.721 and 0.678 for recurrence and treatment outcome prediction respectively.

Computer Aid Diagnosis (CAD) has developed digital pathology with Deep Learning (DL)-based tools to assist pathologists in decision-making. Content-Based Histopathological Image Retrieval (CBHIR) is a novel tool to seek highly correlated patches in terms of similarity in histopathological features. In this work, we proposed two CBHIR approaches on breast (Breast-twins) and skin cancer (Skin-twins) data sets for robust and accurate patch-level retrieval, integrating a custom-built Siamese network as a feature extractor. The proposed Siamese network is able to generalize for unseen images by focusing on the similar histopathological features of the input pairs. The proposed CBHIR approaches are evaluated on the Breast (public) and Skin (private) data sets with top K accuracy. Finding the optimum amount of K is challenging, but also, as much as K increases, the dissimilarity between the query and the returned images increases which might mislead the pathologists. To the best of the author's belief, this paper is tackling this issue for the first time on histopathological images by evaluating the top first retrieved images. The Breast-twins model achieves 70% of the F1score at the top first, which exceeds the other state-of-the-art methods at a higher amount of K such as 5 and 400. Skin-twins overpasses the recently proposed Convolutional Auto Encoder (CAE) by 67%, increasing the precision. Besides, the Skin-twins model tackles the challenges of Spitzoid Tumors of Uncertain Malignant Potential (STUMP) to assist pathologists with retrieving top K images and their corresponding labels. So, this approach can offer a more explainable CAD tool to pathologists in terms of transparency, trustworthiness, or reliability among other characteristics.

In this paper, we present a self-training-based framework for glaucoma grading using OCT B-scans under the presence of domain shift. Particularly, the proposed two-step learning methodology resorts to pseudo-labels generated during the first step to augment the training dataset on the target domain, which is then used to train the final target model. This allows transferring knowledge-domain from the unlabeled data. Additionally, we propose a novel glaucoma-specific backbone which introduces residual and attention modules via skip-connections to refine the embedding features of the latent space. By doing this, our model is capable of improving state-of-the-art from a quantitative and interpretability perspective. The reported results demonstrate that the proposed learning strategy can boost the performance of the model on the target dataset without incurring in additional annotation steps, by using only labels from the source examples. Our model consistently outperforms the baseline by 1-3% across different metrics and bridges the gap with respect to training the model on the labeled target data.