Search

The increasing popularity of attention mechanisms in deep learning algorithms for computer vision and natural language processing made these models attractive to other research domains. In healthcare, there is a strong need for tools that may improve the routines of the clinicians and the patients. Naturally, the use of attention-based algorithms for medical applications occurred smoothly. However, being healthcare a domain that depends on high-stake decisions, the scientific community must ponder if these high-performing algorithms fit the needs of medical applications. With this motto, this paper extensively reviews the use of attention mechanisms in machine learning methods (including Transformers) for several medical applications based on the types of tasks that may integrate several works pipelines of the medical domain. This work distinguishes itself from its predecessors by proposing a critical analysis of the claims and potentialities of attention mechanisms presented in the literature through an experimental case study on medical image classification with three different use cases. These experiments focus on the integrating process of attention mechanisms into established deep learning architectures, the analysis of their predictive power, and a visual assessment of their saliency maps generated by post-hoc explanation methods. This paper concludes with a critical analysis of the claims and potentialities presented in the literature about attention mechanisms and proposes future research lines in medical applications that may benefit from these frameworks.

Abstract Background As of today, cancer is still one of the most prevalent and high-mortality diseases, summing more than 9 million deaths in 2018. This has motivated researchers to study the application of machine learning-based solutions for cancer detection to accelerate its diagnosis and help its prevention. Among several approaches, one is to automatically classify tumor samples through their gene expression analysis. Methods In this work, we aim to distinguish five different types of cancer through RNA-Seq datasets: thyroid, skin, stomach, breast, and lung. To do so, we have adopted a previously described methodology, with which we compare the performance of 3 different autoencoders (AEs) used as a deep neural network weight initialization technique. Our experiments consist in assessing two different approaches when training the classification model — fixing the weights after pre-training the AEs, or allowing fine-tuning of the entire network — and two different strategies for embedding the AEs into the classification network, namely by only importing the encoding layers, or by inserting the complete AE. We then study how varying the number of layers in the first strategy, the AEs latent vector dimension, and the imputation technique in the data preprocessing step impacts the network’s overall classification performance. Finally, with the goal of assessing how well does this pipeline generalize, we apply the same methodology to two additional datasets that include features extracted from images of malaria thin blood smears, and breast masses cell nuclei. We also discard the possibility of overfitting by using held-out test sets in the images datasets. Results The methodology attained good overall results for both RNA-Seq and image extracted data. We outperformed the established baseline for all the considered datasets, achieving an average F1 score of 99.03, 89.95, and 98.84 and an MCC of 0.99, 0.84, and 0.98, for the RNA-Seq (when detecting thyroid cancer), the Malaria, and the Wisconsin Breast Cancer data, respectively. Conclusions We observed that the approach of fine-tuning the weights of the top layers imported from the AE reached higher results, for all the presented experiences, and all the considered datasets. We outperformed all the previous reported results when comparing to the established baselines.

The still prevalent use of paper conformity lists in the automotive industry has a serious negative impact on the performance of quality control inspectors. We propose instead a hybrid quality inspection system, where we combine automated detection with human feedback, to increase worker performance by reducing mental and physical fatigue, and the adaptability and responsiveness of the assembly line to change. The system integrates the hierarchical automatic detection of the non-conforming vehicle parts and information visualization on a wearable device to present the results to the factory worker and obtain human confirmation. Besides designing a novel 3D vehicle generator to create a digital representation of the non conformity list and to collect automatically annotated training data, we apply and aggregate in a novel way state-of-the-art domain adaptation and pseudo labeling methods to our real application scenario, in order to bridge the gap between the labeled data generated by the vehicle generator and the real unlabeled data collected on the factory floor. This methodology allows us to obtain, without any manual annotation of the real dataset, an example-based F1 score of 0.565 in an unconstrained scenario and 0.601 in a fixed camera setup (improvements of 11 and 14.6 percentage points, respectively, over a baseline trained with purely simulated data). Feedback obtained from factory workers highlighted the usefulness of the proposed solution, and showed that a truly hybrid assembly line, where machine and human work in symbiosis, increases both efficiency and accuracy in automotive quality control.

With the increasing adoption of teledermatology, there is a need to improve the automatic organization of medical records, being dermatological image modality a key filter in this process. Although there has been considerable effort in the classification of medical imaging modalities, this has not been in the field of dermatology. Moreover, as various devices are used in teledermatological consultations, image acquisition conditions may differ. In this work, two models (VGG-16 and MobileNetV2) were used to classify dermatological images from the Portuguese National Health System according to their modality. Afterwards, four incremental learning strategies were applied to these models, namely naive, elastic weight consolidation, averaged gradient episodic memory, and experience replay, enabling their adaptation to new conditions while preserving previously acquired knowledge. The evaluation considered catastrophic forgetting, accuracy, and computational cost. The MobileNetV2 trained with the experience replay strategy, with 500 images in memory, achieved a global accuracy of 86.04% with only 0.0344 of forgetting, which is 6.98% less than the second-best strategy. Regarding efficiency, this strategy took 56 s per epoch longer than the baseline and required, on average, 4554 megabytes of RAM during training. Promising results were achieved, proving the effectiveness of the proposed approach.

The scarcity of balanced and annotated datasets has been a recurring problem in medical image analysis. Several researchers have tried to fill this gap employing dataset synthesis with adversarial networks (GANs). Breast magnetic resonance imaging (MRI) provides complex, texture-rich medical images, with the same annotation shortage issues, for which, to the best of our knowledge, no previous work tried synthesizing data. Within this context, our work addresses the problem of synthesizing breast MRI images from corresponding annotations and evaluate the impact of this data augmentation strategy on a semantic segmentation task. We explored variations of image-to-image translation using conditional GANs, namely fitting the generator’s architecture with residual blocks and experimenting with cycle consistency approaches. We studied the impact of these changes on visual verisimilarity and how an U-Net segmentation model is affected by the usage of synthetic data. We achieved sufficiently realistic-looking breast MRI images and maintained a stable segmentation score even when completely replacing the dataset with the synthetic set. Our results were promising, especially when concerning to Pix2PixHD and Residual CycleGAN architectures.

Dermoscopic images allow the detailed examination of subsurface characteristics of the skin, which led to creating several substantial databases of diverse skin lesions. However, the dermoscope is not an easily accessible tool in some regions. A less expensive alternative could be acquiring medium resolution clinical macroscopic images of skin lesions. However, the limited volume of macroscopic images available, especially mobile-acquired, hinders developing a clinical mobile-based deep learning approach. In this work, we present a technique to efficiently utilize the sizable number of dermoscopic images to improve the segmentation capacity of macroscopic skin lesion images. A Cycle-Consistent Adversarial Network is used to translate the image between the two distinct domains created by the different image acquisition devices. A visual inspection was performed on several databases for qualitative evaluation of the results, based on the disappearance and appearance of intrinsic dermoscopic and macroscopic features. Moreover, the Fréchet Inception Distance was used as a quantitative metric. The quantitative segmentation results are demonstrated on the available macroscopic segmentation databases, SMARTSKINS and Dermofit Image Library, yielding test set thresholded Jaccard Index of 85.13% and 74.30%. These results establish a new state-of-the-art performance in the SMARTSKINS database.